A Forecasting Model for Labor Resources in Construction Projects Using Machine Learning

by

Hamidreza Mohammadhosseinzadeh Golabchi

A thesis submitted in partial fulfillment of the requirements for the degree of

Master of Science
in
Construction Engineering and Management

Department of Civil and Environmental Engineering

University of Alberta

© Hamidreza Mohammadhosseinzadeh Golabchi, 2021

Abstract

Considering the high rates of labor resources in construction projects clearly indicates the

importance of appropriate labor resource management methods. Accurate labor resource allocation

is a substantial step towards successful labor resource management. With the recent developments

in the area of artificial intelligence and machine learning, these technologies can potentially be

adopted to develop prediction models. This research aims to combine the benefits of artificial

intelligence and historical data of previous projects to identify the significant factors affecting the

labor resource requirements and to develop an efficient predictive model to analyze and learn from

the past construction projects in order to have a precise estimate of required labor hours for

upcoming projects.

The research involves collecting and analyzing historical data, investigating current industry

practices in labor resource estimation, and implementing machine learning algorithms to predict

required labor hours for various resources in construction projects. Also, this study explores the

key factors impacting the needed labor resources by combining the literature review and industry

practices. Furthermore, this thesis offers a neural network model which can forecast the required

labor hours for a construction work package by utilizing the historical data. The proposed model

provides the estimation of the required labor hours for each day of the work package. The

developed model aids project managers in labor resource allocation in construction projects and

provides them a precise insight to perform a decent labor resource management.

ii
Acknowledgements

First and foremost, I would like to express my sincere gratitude to my advisor Dr. Ahmed Hammad

for his continuous support and consistent encouragement during my studies.

I am also thankful to Maria Al-Hussein for her supportive assistance and many thanks to Trever

Harrington for providing invaluable and professional feedback.

I owe my deepest gratitude to my parents for their never-ending assistance and motivation during

every stage of my life.

Finally, I would like to thank my brilliant brother and his kind wife for their ongoing

encouragement. Without their support, this could not have been possible.

iii
Table of Contents

Chapter 1 Introduction .................................................................................................................................. 1

1.1 Background ................................................................................................................................... 1

1.2 Research Objectives ...................................................................................................................... 2

1.3 Research Methodology ................................................................................................................. 3

1.4 Thesis Organization ...................................................................................................................... 4

Chapter 2 Literature Review ......................................................................................................................... 5

2.1 Human Resource Management (HRM) in Construction ............................................................... 5

2.2 Factors Affecting HR Demand in Construction Projects ............................................................ 11

2.3 Applying Machine Learning Algorithms in Construction Projects ............................................ 14

2.4 Research Gaps ............................................................................................................................. 20

Chapter 3 Methodology for Forecasting Labor Resources ......................................................................... 22

3.1 Introduction ................................................................................................................................. 22

3.2 Significant Key Factors Affecting Required Labor Resources ................................................... 24

3.2.1 Industry Practices in Labor Resource Estimating ............................................................... 24

3.2.2 Selection of Factors............................................................................................................. 26

3.3 Data Acquisition Model for Data Collection .............................................................................. 28

3.3.1 Entity Relationship Diagram (ERD) ................................................................................... 29

3.3.2 Attributes Required for Forecasting Labor Resources ........................................................ 32

3.4 Data Collection Process .............................................................................................................. 33

3.4.1 Essential Attributes for the Forecasting Model ................................................................... 34

3.4.2 Data Sources ....................................................................................................................... 34

3.4.3 Missing Information............................................................................................................ 35

3.4.4 Data Transformation and Linking ....................................................................................... 35

3.4.5 Data Understanding and Visualization................................................................................ 36

3.4.6 Deficiencies in the Collected Dataset ................................................................................. 40

iv
Chapter 4 Development of Forecasting Model for Labor Resources.......................................................... 42

4.1 Introduction ................................................................................................................................. 42

4.2 Data Preprocessing...................................................................................................................... 43

4.2.1. Missing Values.................................................................................................................... 44

4.2.2. Cluster Analysis and Outliers Detection ............................................................................. 44

4.2.3. Nominal Features ................................................................................................................ 47

4.2.4. Normalizing & Data Splitting ............................................................................................. 48

4.3 Feature Selection ......................................................................................................................... 50

4.4 Forecasting Models ..................................................................................................................... 52

4.4.1. Baseline Model ................................................................................................................... 53

4.4.2. Linear Model ....................................................................................................................... 53

4.4.3. Recurrent Neural Network Model ....................................................................................... 54

4.5 Performance Evaluation .............................................................................................................. 57

4.6 Summary ..................................................................................................................................... 59

Chapter 5 Implementation of the Labor Resource Forecasting Model ....................................................... 61

5.1 Introduction ................................................................................................................................. 61

5.2 Application Components............................................................................................................. 64

Chapter 6 Summary, Limitations and Future Work .................................................................................... 72

6.1 Research Summary ..................................................................................................................... 72

6.2 Limitations .................................................................................................................................. 75

6.3 Future Work and Recommendations........................................................................................... 76

References ................................................................................................................................................... 77

Appendix ..................................................................................................................................................... 88

v
List of Tables

Table 1. RNN Model Error Result Before Clustering................................................................................. 44

Table 2. Result of Cluster Analysis ............................................................................................................ 46

Table 3. RNN Model Clusters and Results ................................................................................................. 74

vi
List of Figures

Figure 1. Research Methodology .................................................................................................................. 4

Figure 2. Methodology of Developing the Forecasting Model ................................................................... 23
Figure 3. ERD ............................................................................................................................................. 30
Figure 4. Organization Breakdown Structure Schema ................................................................................ 31
Figure 5. Work Package Attributes ............................................................................................................. 32
Figure 6. Project Attributes ......................................................................................................................... 33
Figure 7. Resource Attributes ..................................................................................................................... 33
Figure 8. Distribution of Work Packages Duration .................................................................................... 37
Figure 9. Distribution of Work Package Types .......................................................................................... 38
Figure 10. Distribution of Resource Categories.......................................................................................... 39
Figure 11. Development of the Forecasting Model .................................................................................... 43
Figure 12. Clustering Analysis Process in RapidMiner .............................................................................. 46
Figure 13. Sample Illustration of Using Dummy Variables ...................................................................... 48
Figure 14. Baseline Model Schema ............................................................................................................ 53
Figure 15. Linear Model Schema................................................................................................................ 54
Figure 16. RNN Model Schema.................................................................................................................. 55
Figure 17. RNN Model Structure ................................................................................................................ 56
Figure 18. MAE of the Forecasting Model ................................................................................................. 58
Figure 19. Interface of the Application ....................................................................................................... 62
Figure 20. Sample Run of the Application ................................................................................................. 63
Figure 21. Initial Version of the Interface................................................................................................... 64
Figure 22. Main Window & Dropdown Buttons (Inputs) ........................................................................... 65
Figure 23. Dropdown Buttons and Lists ..................................................................................................... 66
Figure 24. Main Frame Setup Code ............................................................................................................ 66
Figure 25. Dropdown Buttons Setup Code ................................................................................................. 67
Figure 26. Model's Output Setup Code ....................................................................................................... 68
Figure 27. Output Plotting Setup Code ....................................................................................................... 68
Figure 28. Model's Prediction for Required Labor Resource in a Steel Work Package ............................. 69
Figure 29. Predicted Labor Hours Required for a Piping Work Package ................................................... 70
Figure 30. Actual Labor-Hours Spent for a Piping Work Package............................................................. 71
Figure 31. Cluster Analysis Result ............................................................................................................. 73

vii
Chapter 1 Introduction

1.1 Background

The construction industry has the highest share of employment among all industries in most

countries. Also, labor costs include 30-50% of total project costs in the construction industry

(Heravi and Eslamdoost 2015). Thus, effective labor resource management is essential for the

success of construction projects. Through labor resource management, the demand for labor

resources can be met according to the needs of a project, while achieving a realistic and accurate

estimate of the demand for labor resources. However, effective labor resource estimation is

challenging as construction labor productivity depends on multiple factors.

Currently, labor resource estimating is mostly carried out by project managers utilizing the

common software tools and their own judgements. So, the current practices are mostly relied on

the knowledge and experience of the experts involved in a project which would lead to inaccurate

estimates. Moreover, with the current approach, the potential ability of the historical data collected

from previous projects is ignored.

Also, with the rapid advancements in machine learning algorithms, they could be used as the

powerful tools for developing forecasting models and utilizing the historical data to achieve an

accurate labor resource estimation. Despite the wide adoption of machine learning algorithms in

the past two decades (Gondia et al. 2020), adapting machine learning for labor resource prediction

is still in its infancy (Iwu 2016; Gondia et al. 2020).

Accordingly, the goal of this study is to develop an efficient predictive data analytics model to

analyze and learn from data based on previous construction projects in order to have a precise

estimate of needed labor resource for different types of work packages, which helps construction

1
industries to utilize their collected data and to predict labor resources requirements for their

planned projects precisely.

1.2 Research Objectives

This research aims to develop a machine learning model that would forecast the labor resource

requirements for different work package types. The proposed objective is accomplished by

achieving the following: (1) exploring the current industry practices in estimating the labor

resource requirements; (2) identifying the key factors impacting the labor resource allocation in

projects; (3) evaluating novel machine learning algorithms and their capacity in time series

forecasting; (4) developing a generic prediction model capable to forecast the required daily labor

hours for a given work package.

The academic and industrial contributions of this research are as follows:

Academic contributions:

1. Identifying the current industry practices in labor resource estimation.

2. Exploring the significant factors attributes impacting the labor resource requirements in a

construction project.

3. Developing a forecasting model utilizing machine learning algorithms to predict the labor

resource requirements on the work package level.

Industrial contributions:

1. Proposing a generic data acquisition model to collect project data, providing project

managers a precise insight on their previous projects.

2. A framework to forecast labor resource requirements on work package level utilizing

company’s own previous projects.

2
 3. Providing detailed prediction for the required labor hours instead of limiting the estimation

into a total labor hour for the whole work package.

1.3 Research Methodology

To achieve the research objectives, the following tasks are performed: (1) investigating the factors

that impact the labor resource requirements significantly and assessing their applicability to the

defined problem; (2) exploring the current practices in the construction industry for labor resources

estimating; (3) preparing a data acquisition model and collecting historical data based on it; (4)

implementing feature selection to identify contributing factors and input variables of the model;

(5) developing a forecasting model utilizing machine learning algorithms including linear

regression and recurrent neural network; (6) training the model with the pre-processed data and

evaluating the performance of the algorithms; (7) outputting the predicted labor resource hours for

any new work packages by running the trained model.

3
Identifying Current Developing the Model’s Evaluating &
Practices Forecasting Model Validation

Review of literature Providing data Evaluating model’s

in labor resource acquisition model performance
management

Investigating the Implementing feature Validating model by

impacting key factors selection industry experts

Exploring current Utilizing machine Outputting the

industry practices learning algorithms predicted labor
hours

Figure 1. Research Methodology

1.4 Thesis Organization

This thesis is organized into six chapters as follows. Chapter 2 reviews the literature and novel

methods in labor resource management. It also covers the use of machine learning algorithms in

construction projects. Chapter 3 comprises an overview of the proposed research methodology

with a detailed explanation of the data collection process and the proposed data acquisition model.

Chapter 4 covers the development of the forecasting model utilizing machine learning algorithm

and its implementation in a case study. Chapter 5 elaborates the developed computer application

of the forecasting model. Finally, chapter 6 includes the summary of this research along with the

limitations and suggested future work.

4
Chapter 2 Literature Review

The main objective of this thesis is development of a prediction model for accurate estimation of

Human Resource (HR) in construction projects. This model is founded based on the machine

learning algorithms. The review on prior work touches three different areas: Human Resource

Management (HRM) in construction, factors affecting labor resource demand in construction

projects and applying machine learning algorithms in construction projects.

2.1 Human Resource Management (HRM) in Construction

The construction industry has the highest share of employment among all industries in most

countries (Gurmu and Ongkowijoyo 2020). Also, labor costs include 30-50% of total project costs

in the construction industry (Heravi and Eslamdoost 2015). The construction industry is among

the most dynamic and complex environments (Druker et al. 1996; Wild 2002; Loosemore et al.

2003) and HR-related issues is considered as a major cause of this complexity. Therefore, many

theories and methods of the management of HR have been developed. How to treat, evaluate, and

improve the human-related value of HR, has become an important area of research in the strategic

management in construction industry (Baron 2003; Arnett et al. 2002; Pinker and Larson 2003;

Wang and Yao 1999).

Accordingly, effective HRM is essential for the success of construction projects (Liu and Ballard

2008; Heravi and Eslamdoost 2015; Gurmu and Ongkowijoyo 2020). There has been a widespread

realization that construction industry must improve its HRM performance before it can improve

its overall efficiency, productivity, and cost effectiveness (Hammad 2009; Loosemore et al. 2003).

Effective HRM practices are found to lead to positive organizational outcomes (Becker and

Gerhart 1996) such as turnover (Huselid 1995) and productivity (Katz et al. 1987).

5
In the 1990s a gradual redefinition of the HRM from a personnel advisory role to a performance-

based management activity has happened. However, a precise definition of HRM is difficult since

the issue has been subject to debate within construction management experts and no one has yet

provided a single, authoritative definition of what the concept means. This is mainly due to the

uncertainty of whether the change in title- from personnel management to HRM- has necessarily

coincided with an actual change in the way that organizations manage people or not. Consequently,

there has been still discussion between leading researchers and practitioners as to what the

distinction between the terms is and what this actually means in practice. Somehow, HRM can be

defined as the strategies or policies and practices related to developing the HR of an organization

(Inyang 2011). Due to the important role of HRs on productivity (Jantan et al. 2009) and safety

(Widyanty et al. 2020), there has been an increased trend -as mentioned previously- toward HRM

(Wong et al. 2004; Amrutha and Geetha 2020).

HRM includes the development of knowledge and expertise, and the enhancement of performance

(Garavan and Morley 2006). A powerful HRM system is also recognized as a substantial asset for

construction companies, as an organization productivity is highly correlated with its strategies

(Chen et al. 2003). The development of employees, their eligibilities and the total development of

the organization are the main concerns of HRM.

Researches have demonstrated the positive relationship between HRM practices and

organizational performance to help companies achieve their goals (Antonioli et al. 2013; Buller

and McEvoy 2012; Datta et al. 2005; Huselid 1995; Pfeffer 1998). HRM practices in construction

companies could be defined as plans involved in eliminating HR-related issues in such processes

including recruiting, screening, training, rewarding, and appraising the performance of HRs within

companies (Bowen and Ostroff 2004; Dessler 2012; Huselid 1995). But, in the context of projects

6
and considering the common constraints in time and budget as well as the expectations from

projects' deliverables, common HRM practices could not always be applied (Lim and Mohamed

1999; PMI 2013).

With rapid changes in technology, workers’ needs, current market, financial issues and competitive

environment, planning for HRs have become an important and challenging task for HRM

development. HR planning involves plans for future needs of employees, their required skills,

acquisition of employees, and personnel development (Werther and Davis 1982). Through Human

Resource Planning (HRP), the demand for HRs can be met according to the needs of a project,

while achieving a realistic and accurate estimate of the demand for HRs (Mutua 2019; Reilley

2003). However, effective HRP is challenging as it depends on multiple factors (Heravi and

Eslamdoost 2015). Accordingly, some researchers have adapted Artificial Intelligence (AI) to

address the existing challenges (Jantan et al. 2009). Utilizing AI in construction projects will be

discussed later in this chapter.

The primary concern of HRP is to integrate the strategic and operational requirements of the

project with a workforce equipped to provide the needed services and products (Marchington and

Wilkinson 2002). Some research emphasizes the importance of planning, especially within the

dynamic projects like construction as it can help reduce uncertainty, introduce structure and create

order and action (Laufer et al. 1999). Besides, Turner (2002) represents two primary components

of productive HRP: establishing a strategic HR forecast and preparing HRP framework. The

strategic HR forecasting is a key input to the direction of an organization. HRP is developed to act

as a means to achieve strategic HRM targets, and thus forms the output (Turner 2002).

As some research argue, planning is necessary for increasing competency levels of HRs and it

causes a more efficient management of HRs in an organizational (Tsui 1987; Tabassi et al. 2012).

7
Additionally, planning to improve HR-related issues such as capabilities, collaboration, and team

working skills could eradicate majority of HR-related risks that might appear during projects

(Baiden and Price 2011; Bredin 2008; Campion et al. 1993).

Traditionally, HRP is carried out reactively during the execution of a project, instead of proactively

through consideration of long-term requirements of projects and employees (Raiden et al. 2004).

In construction, project managers usually prefer to allocate their expertise more to solve issues

relating to scheduling, budgeting, risk management, and controlling in projects and mostly

disregard HR-related issues (Scott-Young and Samson 2008; Zwikael and Unger-Aviram 2010).

Besides, there are few studies which inquire influential factors that can lead to a more efficient

HRM within construction projects (Belout and Gauvreau 2004; Huemann 2010). Also, the main

overlooked aspect of HRP frameworks is that they are generally subsets of more comprehensive

project management frameworks and are not specifically considered as an independent framework

developed for the HRs (Davis 2014; Pinto and Prescott 1988). Overall, the studies which

investigate influential factors that could lead to successful HRM are mostly general or limited to

specific aspects such as HR empowerment or HR training (Huemann 2010; Raiden et al. 2004;

Tabassi et al. 2012; Kukenberger et al. 2012; Pant and Baroudi 2008).

Accordingly, comprehensive HRP frameworks should be developed and modified to indicate main

aspects of HRM in projects including assessment, training, recruitment, and development (Bourne

and Walker 2005; Partington et al. 2005). Some studies tried to fill these gaps by trying new

approaches for HRP framework (Baloh and Desouza 2012; Henver 2007). For example, Pournader

et al. (2014) investigated a three-step design science approach consisting of rigor, relevance, and

design cycles. Rigor and relevance cycles aim at proposing the initial HRP framework and

conducting a qualitative exploratory study of two construction-engineering companies to establish

8
the validity of the framework, respectively. At the final step, the design cycle evaluates the

applicability of the developed framework. The proposed framework is investigated by

quantitatively testing through conducting a survey of 110 construction industry experts.

Traditionally, construction projects consist of several stakeholders including end users, promoters,

project designers, project team, and work force. Hence, the wide variety of HRs involved in

projects with different expectations would require their needs to be reflected on project's

deliverables as well (Cleland and Ireland 2006; Davis 2014; Ballesteros Pérez et al. 2010). Thus,

despite the fact that developing HRP frameworks for large organizations are common practices in

the research literature (Becker and Huselid 2006; Wright and Boswell 2002), these frameworks

usually do not consider HR issues as unique features of projects' environment within organizations

(Belout and Gauvreau 2004; Huemann 2010).

Nevertheless, a number of attempts have been made to present HRP frameworks and to highlight

the impacts that HRM have on construction projects. For instance, Tsui and Milkovich (1987)

explored HRM through various aspects such as labor resource staffing, compensation, and training.

Turner and Müller (2005) studied the fundamental competencies of project managers to lead

projects towards success. Belout and Gauvreau (2004) compared the overall impact of HRs as

personnel on the different aspects of a project for its successful implementation. More recently,

Davis (2014) identified a limited number of HR-related factors that affect viewpoint of different

parts involved in a project regarding project execution. There are also similar studies that point out

benchmarks of project's success, which generally include HRM success factors (Verburg et al.

2012; Pinto and Slevin 1988). However, considerable amounts of these studies have seldom

considered the applications of the HRM and its significant role in success of projects. Moreover,

as an international project management standard, PMBOK (PMI 2013) introduces HRM in four

9
consecutive sections including ‘Plan Human Resource Management’, ‘Acquire Project Team’,

‘Develop Project Team’, and ‘Manage Project Team’. However, it seems PMBOK introduces

HRM as a rather general framework in projects which could profoundly affect HRM practices in

projects (Pournader et al. 2014).

In order to make the construction projects more manageable and approachable the projects can be

broken-down into small work packages and each work package contains information related to the

deliverables such as durations, resources, risks, etc. (PMI 2013). In the context of construction

projects, the Work Breakdown Structure (WBS) is the tool that utilizes this technique. According

to PMI, WBS is “deliverable oriented hierarchical decomposition of the work to be executed by

the project team.”. WBS is used for many different purposes. Initially, it is used as planning tool

to define and organize scope with deliverables. PMI defines deliverable as any unique product,

result or service which needs to be produced to complete a task, phase, or project. (PMI 2013).

WBS also can be used as a source for accurate scheduling and resource estimating. In this study

to achieve a better prediction for resources, specifically human resources, the developed model

will be applied for each work package. In other words, each construction project is initially broken-

down into smaller work packages and consequently the prediction will be done for work packages.

There are two types of WBS: Deliverable-Based and Phase-Based (PMI 2013). The main

difference between these two approaches is the elements identified in the initial level of

breakdown. In a Deliverable-Based approach the relationship between the deliverables and the

scope is clearly elaborated. On the other hand, in Phase-Based approach the deliverables are

demonstrated for each phase of the construction project.

10
2.2 Factors Affecting HR Demand in Construction Projects

Achieving an efficient HRP requires identifying the factors that have an effect on HR within

construction projects. The simulated models reveal that the HR demand in construction projects is

based on multiple different factors like project size and project type. The larger the size of the

project is; the more HR is required for a particular project type. There is a wide variety of studies

devoting to these influential factors and evaluating their impacts on HR demand in construction

projects. For instance, Chan et al. (2003) showed the strong relationship between HR demand and

project size in an analysis of 123 construction projects. Some of the most important factors are

discussed as follows:

Type and Size of the Project

The HR demand for a construction project is closely related to the type of project as different

construction projects tend to have a different product mix and fixed cost structure (Agapiou et al.

1995; Chan et al. 2002). For example, some trades such as plasterers and more technical skilled

workers are closely associated with new housing work, whereas scaffolders have more

employment opportunities from general repair and maintenance activities (Briscoe and Wilson

1993). The combination of skills also changes significantly when construction shifts from piling

work to the construction of the superstructures. For instance, building a rural traditional house

certainly requires more physical labor but less plant than a prefabricated building. Obviously,

project size and type are important factors that dictate the extent to which specialized skills are

practiced in the construction industry (Persad et al. 1995).

However, some researchers have identified a number of additional factors which have an impact

on the HR requirements of construction projects. These include construction method, degree of

11
mechanization, project complexity, management attributes, expenditures on electrical, mechanical

services, construction output and wage level.

Project Delivery & Construction Method

The construction method of an individual project and its delivery method impact the needed labor

resources and combination of skills (Lemessany and Clapp 1978). For example, a residential block

with traditional brick walls requires significantly more HR to execute than those, which were built

using industrialized systems of construction concrete based. The increasing use of prefabrication,

production activities off-site, and the use of other construction methods have caused a reduction

in the demand for traditional craft skills like bricklaying, plastering and carpentry, but an increase

in prefabricated elements erectors (Agapiou et al. 1995; Tang et al. 1990). Accordingly, the

growing use of prefabricated components results in over 40% reduction in the consumption of HR

in construction sites (Tam et al. 2006).

Moreover, the utilization of automation and mechanized methods also significantly influences the

required labor resources (Ehrenberg and Smith 2003). In general, the more the automated

equipment is utilized, the less the labor resources are required since automation tends to be labor

saving (McConnell et al. 2003).

Project Complexity

Another factor affecting HR demand at projects is the complexity of the construction project

(Ganesan et al. 1996). Gidado and Millar (1992) defined complexity as a significant factor

impacting the labor demand on site including: technical complexity of the activities, amount of the

overall and interdependencies in construction stages, project organization, site condition, and

uncertainties of the work on site. For example, the design of the Bank of China in Hong Kong

12
supports the idea that structural design can potentially bring in resource savings. The total required

steel was nearly half of a typical building with the same. The reduction on the usage of steel lead

to less labor resource requirements for fixing and alignment of frames. Similarly, the design and

use of modern hydrant systems decreases wiring requirements which leads to less associated labor

service requirements (Fairweather 1986).

Handy (1985) introduces project size as a single factor in investigating the appropriate construction

team organization. Wong et al. (2004) believes four attributes including overall technological

complexity of overall project characteristics, site physical site condition, buildability level and

complexity of coordination works are the important factors which might have an impact on the

project HR demand.

Management Attributes

HR requirements are also affected by contractor’s management skills such as planning, organizing,

and controlling (Wong et al. 2003; Gould 2002). Appropriate coordination and utilization of

resources especially labors on sites would lead to reduction in HR requirements (Ganesan et al.

1996). Enhanced HRM can lead to HRs saving in projects and better interfacing between different

trades, such as electrical and mechanical trades. Detailed precise planning of site work could also

cause reduction in labor resource requirements. For instance, in laying pipes and conduits, last-

minute changes in design often result in abortive HR (Gruneberg 1997).

13
2.3 Applying Machine Learning Algorithms in Construction Projects

Construction projects comprises a number of different activities, which relate to and impact upon

one another and are affected by various uncertainties, such as weather, geological characteristics

and humans. Professional construction management is necessary to accomplish construction

objectives efficiently and is essential to project success (Bush 1973). Due to uncertainties and the

changing nature of the construction industry, practical construction management problems are

complex and hard-to-predict (Li 1996). In fulfilment of these problems different tools have been

successfully developed and applied in construction management. Naturally, humans are able to

learn and can process complex problems even in the presence of uncertainties and insufficient

information. Following the process of human inference offers an effective approach for solving

construction management problems. Artificial intelligence (AI) relates to computer system designs

that handle and attempt to resolve problems intelligently by emulating processes inside the human

brain. As AI technology enhances the ability of computer programs to handle tasks for which

humans are currently still better at handling, employing AI paradigms is appropriate in efforts to

solve construction management problems (Haykin 2007; Tommelein et al. 1992). Various

scientific and engineering fields have been paying increasing attention in recent years to different

artificial intelligence (AI) paradigms.

Machine learning -a subset of AI- is considered as one of the top technologies used widely in

different industries (Muizz et al. 2020). Machine learning is used for data modelling i.e.,

developing mathematical abstractions of data that can be used by computers to provide accurate

prediction. Despite the wide adoption of machine learning algorithms in the past two decades

(Gondia et al. 2020), adapting machine learning for construction management is still in its infancy

(Jantan et al. 2009; Iwu 2016; Gondia et al. 2020).

14
In the construction industry, studies have been performed to make various predictions utilizing

machine learning. For example, Tixier et al. (2016) performed random forest analysis and

stochastic gradient tree boosting with 4400 validated datasets, and Gerassis et al. (2017) analyzed

six years of accident data. They used a Bayesian network to predict the probability of the accident

type for bank-related accidents. Amiri et al. (2016) analyzed five years of data in Iran using

multiple-correspondence analysis, decision tree analysis, ensembles of decision tree, and the

association rules method. Alizadeh et al. (2015) calculated the conditional probability of severe

and fatal injury between the parameters of age, marital status, career, accident experience, and

accident severity employing Bayesian theory. Furthermore, they tried to apply the achieved results

in workers’ training process in order to mitigate accidents and improve their insight about the

perception of probable risks on site. Lastly, Chiang et al. (2018) conducted a cluster analysis of

fatal accidents in Hong Kong. With the rapid advancement of machine learning technology, more

approaches have aimed to use machine learning to address the challenges associated with HRM

(Jantan et al. 2009; Heravi and Eslamdoost 2015; Xie 2020). Moreover, some studies have shown

that using machine learning algorithms improves the precision of the HR-related process (Jantan

et al.2009; Wang et al. 2017).

In this section, a summary of common state-of-the-art machine learning algorithms mostly used in

construction (KNN, ANN and Random Forest) is presented. Detailed mathematical descriptions

of these techniques may be found in relevant references (Wauters and Vanhoucke 2017; Wang et

al. 2016; Haykin 1999).

K-Nearest Neighbor (KNN)

K-Nearest Neighbor algorithm is one of the most fundamental and simple classification methods.

The KNN algorithm is a simple algorithm that is based on predicting new records through

15
similarity measures. KNN classification was originally developed to deal with unknown reliable

parametric estimates of probability density during discriminant analysis. Many studies have

utilized this algorithm in different research areas as solutions by learning methods, mapping, and

recognition. (Dang et al. 2005; Franco et al. 2001; Lee and Scholz 2006). The KNN algorithm is a

powerful tool in dealing with classification and learning from massive datasets (Rosa et al. 2003).

In KNN, k refers to the number of neighbors included in the majority of the voting process (Gou

et al. 2003). More specifically, KNN classifier needs a metric d and a positive integer k value

(Kubat and Cooperson 2001). When a new input requires classification, the distance between the

new data and training records is being calculated. Based on the specified threshold for the number

of neighbors (which is k), k nearest records with the least distances are detected and selected. The

class with more samples would be selected as the class of the new input. For example, if k is equal

to three, the prediction of a new record will be calculated from the majority of votes from the three

nearest neighbors.

As a result, the only parameter that needs to be tuned is the value of k. A smaller value of k can

bias the model towards outliers and a larger value of k can make the modeling process

computationally expensive (Gou et al. 2003). The optimal value of k resulting in the best

performance could be identified through trial and error and there are also existing studies

suggesting different formulas for the optimal value of k (Zhu et al. 2010; Lall and Sharma 1996).

In KNN algorithm, different methods can be used for computing the distance. Euclidean distance

metric is a simple and easy method for calculating distances in multidimensional input space. This

method is widely used and can yield competitive results even compared to the most complex

machine learning methods (Song et al. 2007; Duda et al. 2001).

16
Artificial Neural Network (ANN)

Networks are a smart technique for taking a complex system into simpler subsets to achieve clear

understanding (Wu and Chan 2009). A set of nodes and connections between nodes are main

components of networks (Muizz et al. 2020). The nodes are considered as computational units of

networks and the connections establish the information flow between these nodes (Haykin 1999).

In ANN method, the nodes are known as artificial neurons which is a computational model

originated from the natural neurons. In the artificial neurons, inputs are multiplied by assigned

weights and then calculated by a predefined activation function (Patterson 1996). ANNs are

comprised of artificial neurons (Gershenson 2003). In other words, ANNs were built to process

information in a manner similar to the human brain and consist of a set of interconnected input and

output units where each link has an associated weight. The ANN technique enables modeling large,

complex problems that involve many interrelated variables (Mourya and Gupta 2012). ANN

method has a superior ability in prediction, pattern recognition, data compression, and decision-

making (Chukwu and Adepoju 2012; Paliwal and Kumar 2009).

Recently, there are many different developed models of ANNs. The differences are mainly in the

activation functions, the hybrid models, the accepted values, and the learning algorithms (Wu and

Chan 2009). For example, recurrent neural network is one of the most state-of-art models of ANNs

(Hibat-Allah et al. 2020). The recurrent neural network is commonly used in speech recognition

and natural language processing, and modeling sequence data (Zaremba et al. 2014). The algorithm

has the ability to memorize the sequential characteristics of the data and use patterns to forecast

the probable scenario (Apaydin et al. 2020). The recurrent neural network allows previous outputs

to be considered as inputs while having hidden layers (Apaydin et al. 2020).

17
ANN models usually require prolonged training time and interpreting the meaning of the

calculated weights of the nodes can be difficult (Han et al. 2011). ANNs can be used in both

classification and regression problems (Mourya and Gupta 2012).

Random Forest

The random forest algorithm is considered as one of the most precise forecasting methods (Wang

et al. 2016). Random forest is an ensemble method for both classification and regression consisting

of decision tress. Each tree consists of a root node that is divided into branches based on all possible

outputs. The splitting is repeated for each branch until reaching a node where all instances have

the same classification (Witten et al, 2011).

The main objective of the random forest algorithm is to provide a robust prediction model that has

a better performance and is less sensitive to overfitting, through averaging several decision trees

which might individually suffer from high variance (Bonaccorso 2018). The splitting approach in

random forest is based on a medium level of randomness as opposed to a single decision tree

(Probst et al. 2019). A randomly selected set of features in each tree is used to achieve the threshold

that best splits the data. Thus, there are several trees that are trained, and each one would perform

a different prediction (Probst et al. 2019). Moreover, the trees which have the best performance in

a portion of the sample space, might provide inaccurate estimates in other portions (Pedregosa et

al. 2011). Finally, prediction is made by aggregating (majority vote or averaging) the predictions

of the ensemble (Wang et al. 2016).

The random forest algorithm represents many advantages. The main advantages of the Random

Forest algorithm include the simplicity of the generated rules and the learning and classification

process, while not being limited to numerical or categorical data (Gorunescu 2011). It runs

18
efficiently on massive datasets. The algorithm is also not sensitive to noise or over-fitting. It is

capable to handle large number of inputs and meanwhile has few parameters compared to the other

machine learning algorithm like ANN.

Due to its capabilities, there has been recently a trend towards utilizing random forest algorithm.

Consequently, there are studies in construction industry in which researchers try to use this

algorithm as the potential solution of various problems such as predicting occupational accidents,

risk delay in construction projects, and project costs (Yaseen et al. 2020; Huang & Hsieh, 2020;

Meharie and Shaik 2020)

19
2.4 Research Gaps

Labor resource has always been the most critical part of the construction resource pool. Effective

management of resources especially labors is a complex issue which can be the guiding factor for

the success or failure of any project. Although resource allocation has been widely studied by

construction researchers, but project managers are still struggling with resource management

difficulties. One of the key solutions are the forecasting models which can potentially provide

practical answers for resource allocation problems.

The developed forecasting models are mainly focused on predicting the total labor demand for the

entire project based on the general specifications of the project. Current models mostly consider

the project as a single entity without dividing it into smaller packages. This approach provides

limited insight for the required labor resources. Although these models can provide project

managers with a general idea of the required resources, but they do not include many details about

the predicted result. The point is that having a rough estimation about labor demand for the whole

project without knowing the distribution of the demand over the time would not help project

managers efficiently. On the other hand, there are some models which does predictions at the

activity level. They assume each construction project as a group of sequential activities and

forecast labor requirements for each of them. But in the early stages of a project there might not

be sufficient details regarding project activities; therefore, making predictions for activities could

lead to high errors and inappropriate resource allocating. Accordingly, there is a need for a

predictive model which can provide estimations in a sufficient level such as work package level.

The current research aims to combine the current industry practices and literature to develop a

forecasting model which is capable to predict the daily labor resource requirements for

construction projects at the work package level. Having an insight for the demands at the work

20
package level is more efficient and also it is more practical than making rough estimations for the

entire project. This approach also would not need detailed information about every planned activity

during the project. The purpose is to have a generic model that provides project managers reliable

results. This goal is fulfilled by: (1) proposing a data acquisition model to aid project managers in

collecting the project information in a sufficient way which would lead to more accurate

prediction; (2) developing a forecasting model utilizing machine learning algorithms to forecast

the required labor resources over the project duration.

Some of the similar predictive models suggest linear regression methods, while others utilize non-

linear approach to forecast the labor resources. This research explored different machine learning

algorithms, including linear regression and neural networks, due to their capabilities in dealing

with massive datasets and providing precise predictions.

21
Chapter 3 Methodology for Forecasting Labor Resources

3.1 Introduction

The objective of this research is to forecast the labor resources requirements for the work packages

of an upcoming project utilizing historical project data and to develop a framework for collecting

significant factors that would help in predicting the distribution of the required hours during the

project. The research methodology utilizes the proposed approach by Ghazal & Hammad (2020)

to develop a reliable forecasting model that can be used by experts, as shown in Figure 2. This

framework is comprised of multiple steps, including understanding industry practices and

techniques, investigating researchers’ approach in forecasting required labor resources at the work

package level, and exploring the key factors affecting the required hours for different labor

resources. In the next step, a data acquisition model is proposed for collecting project information

properly for forecasting labor resource requirements on a work package level. Then, historical data

is collected according to the findings of the literature review and industry practices. After data

cleaning and pre-processing, machine learning algorithms are utilized to identify the significant

factors and develop a model that can forecast the distribution of the required hours for different

types of labor resources. At the last step, the developed model is validated by the industry experts.

22
 Raw Data Predicted
Collection Output

Literature Data Acquisition Forecasting

Review Model Model

Model
Training

Current Industry Data Cleaning & Evaluating &

Selecting Best
Practices Pre-processing
Algorithm
Initial Input
Variables
Final Input
Variables
Feature Selection Machine Learning
Algorithms

Linear Regression

Recurrent Neural
Network

Figure 2. Methodology of Developing the Forecasting Model

The objective is to explore the project features that affect the labor requirements for different work

packages of a construction project. Besides, the developed predictive model is capable to find the

weights of these factors and interpret the results. It is noteworthy that the result needs to be

reasonable and comprehensible to construction experts especially project managers, so it can

provide them reliable support in making resource allocation and workforce planning.

The focus in this chapter is on: (1) investigating resource allocation methodologies adopted by

industry practitioners; (2) identifying the significant project factors that have an influential impact

23
on predicting resource requirements through analyzing the previous research and discussions with

experienced project managers; (3) developing a data acquisition model to ease proper data

collecting for project managers; (4) collecting historical project data to be utilized in the

development of the forecasting models; (5) data cleaning and pre-processing to prepare it for

further exploration, analysis, and data mining; (6) describing the factors in the dataset and

representing the collected data through graphs and charts to elaborate the distribution of factors.

In the next chapter, the process of forecasting model development is discussed in detail, including

the data preparation approach, feature selection, defining the model inputs, and machine learning

algorithms used to predict the hours required for labor resources at work package level for an

upcoming construction project.

3.2 Significant Key Factors Affecting Required Labor Resources

Estimating the required workforce has always been one of the project managers’ concerns. There

are different methods adopted by the industry practitioners and researchers to estimate the labor

resources required for a project. One of the objectives of this research is to explore the potential

factors impacting the resource requirements on a work package level and to determine the

significant factors that can be used to forecast labor resource requirements. Additionally, further

exploration is performed to outline the current industry practices. The investigation is conducted

to identify the project characteristics that need to be considered for estimation of the required labor

hours for different work packages of a construction project.

3.2.1 Industry Practices in Labor Resource Estimating

24
Exploring the current industry practices is a critical step in detecting the key attributes required for

prediction. The target is to implement an investigation approach for impacting factors considered

by industry experts to estimate the labor resources required to complete a work package.

Informative discussions with experienced project managers working in the largest construction

companies lead to a clear understanding about the current resource allocation methodologies in the

industry. Typically, each organization has its own way of identifying the required staff. They adopt

both qualitative and quantitative techniques, including heuristic rules based on their experienced

managers, regression models from previous projects, and their own framework to rank projects

based on multiple project characteristics.

Some common resource estimation methods used by industry organizations include:

1. Expert judgment: Construction organizations usually rely highly on their experts’

evaluations. Companies bring in experienced specialist who have done this sort of work

previously and getting their opinions on what resources are needed.

2. Project management software: The developed software such as Microsoft Project often

have features designed to aid project managers to estimate resource requirements and

constraints and find the best combination of assignments for the upcoming project.

3. Bottom-up estimating: Breaking down complex tasks into smaller activities and working

out the resource estimating for each activity. Actually, it is the process of estimating

individual activity resource need and then adding them up together to a total estimate. This

method is very common among construction companies due to its accuracy and simplicity.

However, it takes a considerable amount of time to perform bottom-up estimating as every

activity must be assessed and estimated precisely to be included in the calculation.

25
 4. Published estimating data: Many project managers in construction industry utilize such

data to figure out the resource requirements. They rely on articles, books, or journals for

analysis.

In the process of labor resource allocation, various techniques including qualitative or quantitative

are adopted. However, each of these frameworks have limitations that impact the accuracy and

lead the practitioners to inaccurate estimation and refrain from using their methodologies. The

main drawbacks can be summarized as follows: solely relying on project managers’ knowledge

and experience, not analyzing the actual project values compared to the estimated values and the

allocation methodologies cannot be generalized across departments and project types.

3.2.2 Selection of Factors

The key attributes impacting labor resource demand are identified by exploring the literature

review and determining the factors applied by industry experts through understanding their

methods in estimating the required workforce. One of the objectives of the current research is to

collect the significant key features from historical data to forecast the labor hours of different

resources required for work packages of a given construction project.

As mentioned in the literature review chapter, there are a few prediction models developed by

researchers to estimate the labor required for a construction project. These models consider various

factors in their developed forecasting models such as the project type, cost, and floor area while

ignored other factors such as complexity and project delivery method (Elkholosy 2020). The

current models are mostly able do predictions only for the “total” labor hours required for a project

while this research tries to develop a model which is capable to predict the required labor hours

for each time step (day or week) during the project. The most frequent factors impacting the

project’s labor resource requirements considered by the researchers in developing their models

26
include: (1) project type (Yang & Kim, 2019; Chen et al., 2008; Bell & Brandenburg, 2003); (2)

cost (Wong et al., 2008; Chen et al., 2008; Bell & Brandenburg, 2003); (3) work type (Yang &

Kim, 2019; Othman et al., 2011); (4) floor area (Chen et al., 2008); (5) project duration (Chen et

al., 2008); (6) site condition (Yang & Kim, 2019).

Additionally, some factors that are mentioned by researchers cannot be generalized to all types of

projects. For example, the floor area can only be considered for estimating buildings projects.

After investigations in the literature review and discussions with industry experts the following

factors were selected to be explored in this study:

1. Work package type: which describes type of the work defined for the work package such

as welding, concrete, piping, cut & fill, etc.

2. Resource category: which indicates the type of labor resource used for the work package

such as carpenter, ironworker, pipefitter, etc.

3. Duration: which elaborates the total time (day or week) in which a specific type of labor

resource is involved in a work package.

4. Key quantity: representing the total amount of a unit of measurement to complete a work

package which can be in different units depending on work package type. For example,

500 cubic meters for a concrete work package or 100 square meters for a piping work

package.

5. The cost: which indicates the total dollar value spent to complete a given work package.

6. The complexity: which elaborates the technical and technological complexity of

completing a work package.

7. Project phase/sub-phase: which represents the stage of the project such as planning,

procurement, construction, etc.

27
Some of factors that might have an impact on the labor resource requirements are not considered

in the case study of this research due to unavailability or confidentiality. For example, considering

the impact of budget on the required labor hours is disregarded since the industry partner did not

provide this information as it was considered confidential. Moreover, the complexity of a work

package, which could be a factor affecting the needed labor hours, is difficult to collect as

contractors do not include such information in their records and each contractor might have its

own definition of complexity.

Thus, a total of 6 factors was used as an input to the forecasting models in this research. These

factors include: work package type, resource category, duration, key quantity, work package cost,

and complexity. The selected factors are then investigated by experienced project managers to

ensure that these factors are correctly chosen. Also, the project managers are asked to provide

feedback if there are any overlooked factors that need to be considered.

3.3 Data Acquisition Model for Data Collection

Appropriate data collecting is a vital and critical step for implementing the information and

analyzing it. Utilizing the historical information of previous projects can improve the overall

performance of the organization by increasing the efficiency and reducing the expenses (Elkholosy

2020). This will help the organizations to have a competitive edge in the market. Recently, the

construction industry has started to move towards this path to be able to do predictive analyses,

productivity improvements, proactive decisions regarding safety, etc. Therefore, obtaining good

quality data is crucial for the organization success.

Since most of the construction companies have their own tracking systems, each company collects

the data in a different way. For example, contractors mostly focus on tracking labor hours without

28
considering other aspects of the projects. On the other hand, owners usually tracking the progress

of the tasks and the costs rather than the spent labor hours. This variance in the process of data

collection would cause difficulties in gathering datasets from various companies and implementing

them into a generic model. Therefore, this study tries to propose a generic data acquisition model

to provide a tracking system which is enabled to collect data from multiple projects. Then a general

forecasting model is developed which can also be utilized by different construction companies

with various tracking systems.

By exploring the business process of the construction contractor in the case study, it is noticed that

their tracking systems is well organized and accurate due to the direct relation between the spent

hours and expenses. Therefore, missing information is not a big concern while the data is pulled

from their tracking systems. As the other contractors might not collected their data in such efficient

way, a data acquisition model is prepared to aid in collecting the required project data for

estimating the labor resources in a structured way. The model will improve the data collection and

the resource allocation processes in an organization.

The provided approach facilitates the process of the storing data and allows the organizations to

benefit more from their stored data and to track their project data properly in a generic way. The

outcome will support the companies in utilizing the labor resources forecasting model by collecting

high quality data. Also, it will help them in performing future analysis that could potentially

provide them with precise insight to take corrective actions and make proactive decisions. An

Entity Relationship Diagram (ERD) is developed to illustrate the different entities involved in the

framework and the relationships between these elements.

3.3.1 Entity Relationship Diagram (ERD)

29
The initial step to achieve a well-organized acquisition model is to have a good understanding of

the attributes and their connections. The goal of developing the ERD, illustrated in Figure 3, is to

represent the relationships between the different entities that need to be tracked and stored in the

database. The purpose is to track the labor hours spent by resources for each project work package

since the required resources varies from one work package to another. Also, the quantity of

resources alternates during the project, so it is far from reality to consider the hours required for a

project is spent equally during the days of the project. The developed framework helps the

contractors in collecting the required project data for labor resource forecasting.

Figure 3. Entity Relationship Diagram

The framework is designed for the contractors to collect the project data in a proper way; hence, a

contractor entity is included in the ERD that the other entities are built on. The designed entities

30
in the ERD include multiple attributes that is represented in the following section. Each

organization might have multiple business units or divisions which include various departments

(Figure 4). The model provides three levels of the breakdown structure to reach the level at which

the projects are managed. The department have specific types of resources with the quantities that

indicate the available workforce. The resource entity involves different labor types such as welder,

carpenter, pipefitter, project manager, etc.

Organization Organization #1

Division Engineering Procurement

Department Design Construction Supply

Figure 4. Organization Breakdown Structure Schema

According to the breakdown structure, a portfolio is comprised of multiple programs that would

help fulfil organizational achievements through strategic business objectives along the way

(Martinsuo & Lehtonen 2007). A group of related projects create the program. the program’s goals

cannot be completed without combining these projects (Blichfeldt & Eskerod 2008). Portfolio

management is necessary for the success of the organization. It allows the company to assess,

prioritize, and choose projects compatible with the overall strategy and assign resources to projects

evaluating the priorities and policies (Elkholosy 2020; Meskendahl 2010).

31
3.3.2 Attributes Required for Forecasting Labor Resources

Each project is comprised of multiple work packages. Moreover, there are one or more contracts

in a project that provide various services, including consulting, construction, soil investigation,

etc. Each work package requires specific types of resources to execute its scope and deliver its

requirements. Therefore, the forecasting of labor resource requirements should be performed on

the work package level when the organizations store the data.

Accurate labor resource prediction cannot be conducted for each work package without

considering work package and resource features. Accordingly, the project and required resources

are the essential entities in this research that need investigation since the forecasting of the

workforce requirements is reliant on their features. The work package and resource characteristics

are illustrated in the ERD with the attributes that need to be collected in Figure 5, 6 & 7. Besides

to the forecasting model, the collected data can provide the contractors more benefits such as

further analysis for the inherent trends and making decisions.

Name KeyQuantity UnitOfMeasure

Type Completion
WORKPACKAGE

Duration Category

Budget ActualCost StartDate FinishDate

Figure 5. Work Package Attributes

32
 ProjectID Location Program

ProjectName Portfolio
PROJECT

Owner
Complexity

DeliveryMethod ProjectType

Figure 6. Project Attributes

ResourceName ResourceQuantity

Organization Unit
RESOURCE
Discipline

ResourceDescription ResourceStatus

Figure 7. Resource Attributes

3.4 Data Collection Process

The high quality data for training plays an important role in the accuracy and reliability of the

developed forecasting model. In this study, historical information consisting of different work

packages is collected from contractors who were involved in huge construction projects taking

place in 2019 and 2020.

Collecting the data is the initial step to utilize it in the forecasting models to estimate labor resource

resource requirements. The information that needs to be collected is proposed in the literature

33
review and the industry practices. In this research, the purpose is to gather historical data on the

work package level. In this study, the historical data is gathered through the spreadsheets, received

from the industry partner, which are originally derived from company’s Enterprise Resource

Planning (ERP) systems.

3.4.1 Essential Attributes for the Forecasting Model

The critical factors that affect the resource demand are determined through reviewing the literature

and tracking the industrial practices of construction contractors. The followings are the essential

attributes of a project and its work packages: work package duration, and complexity; work

package type (concrete, excavation, steel, etc.), project cost, resource category (carpenter,

pipefitter, ironworker, etc.), and key quantity. In this study, the data collection is focused on

collecting the attributes for the historical projects which are completed. The labor resource hours

spent on the work packages are collected and used to train the forecasting model.

3.4.2 Data Sources

Each contractor has its own way of collecting project data. Some might use their own developed

collecting systems, or some may utilize the ERP systems such as Oracle.

The industry partner involved in this study has developed its own tracking system. The tracking

system includes different spreadsheets each used for collecting specific data. The hours spent on

each task is collected on a daily basis. Each task has a unique function code which is the key to

link spreadsheets to each other. These report spreadsheets include the task description, the break

down structure for activities, function description, employee type, work date, planned key quantity,

actual key quantity, and hours spent on each day.

34
3.4.3 Missing Information

Although the contractors usually track the tasks and their spent hours accurately, but there are

some attributes which contractors mostly do not collect such as complexity or location. Moreover,

they might not use the same proper format for collecting data from all projects. For example,

without having a well-defined format, naming different tasks or resources can get complicated and

confusing as each project manager might use his/her own way of naming various tasks or

resources. Having a predefined data acquisition system can prevent all these problems and

provides a clean and well-organized dataset which can really benefit the company in accurate

analysis, quick spotting any possible issue and future planning.

The industry partner involved in this study has a powerful tracking system which is developed

within the company and has been utilized and improved during years. As a result, there were only

limited missing values in their collected data. But as mentioned before, there are some attributes

which the contractor usually does not include in their tracking systems. The missing values are

removed from the training dataset through outlier detection process which will be discussed in

detail later.

3.4.4 Data Transformation and Linking

As there might be different timesheets for collecting various aspects of the project and most of

these attributes may not be necessary for the research purpose, developing a well-defined database

is required. This would help to collect the needed data in a structured manner. Hence, the collected

data from different systems is collected in an organized database ready for importing into the

forecasting models.

The historical data collected from the contractor was compromised of different spreadsheets. Each

of these spreadsheets covered a specific aspect of the project. For example, one was the timesheet

35
of hours spent on each day describing the tasks and resources used on it. The contractor has a

predefined hierarchical list of tasks and resources involved in a project calling each level “parental

code”. There are 3 levels of “parental code” defining the break down structure of each task. For

example, “Deepen Ditch along West Perimeter Road” is a task done during the project and is the

lowest break down level called “parental code 3”. The upper level of this task “parental code 2” is

defined as “Site Drainage” and the highest level “parental code 1” is called “Sirte Drainage WO1

– Athabasca”. Another spreadsheet included the actual and planned key quantities for the project

tasks. Each task has a unique “functional code” which act as the key attribute for linking different

spreadsheets together.

There were various labor resources who worked on the work packages. Therefore, it was essential

to aggregate the similar types of resources as one resource type, because then there would be too

many resource types and the forecasting models were not able to perform an acceptable prediction.

3.4.5 Data Understanding and Visualization

6 key attributes which significantly impact labor resource requirements are identified. They are

used as inputs into the forecasting models. The following section briefly explains these factors

with the illustrations of their distribution in the collected dataset.

3.4.6.1 Work Package Duration

The work package duration is in days, and it represents the time between start and finish date

represented in the tracking system. The purpose of applying this factor is to indicate the effect of

the duration on the labor resource hours required to finish the work package.

36
 60

50

40

30

20

10

Duration (day)

Figure 8. Distribution of Work Packages Duration

3.4.6.2 Work Package Type

Work package type represents the category of the work that is done such as bolt up, cut & fill, or

piping. In the case study there are 11 types of work packages including: Concrete, Culvert, Cut &

Fill, Excavation, General, Piping, Rework, Services, Site Drainage, Steel, and Superintend.

37
 Frequency of Different Work Packages in Collected Dataset
120

100

80

60

40

20

Figure 9. Distribution of Work Package Types

3.4.6.3 Work Package Complexity

The complexity of the work package is defined through a 5-level scale to represent the technical

complexities and the uncertainties. The scale is designed to measure the required effort for

completion of work packages with different level of complexities. As mentioned before, there are

some attributed which the contractors do not include in their tracking systems or there might not

be defined way to measure such ones. The complexity is among those attributes which is not

included in their dataset. Accordingly, this attribute is not included in the case study due to

unavailability.

3.4.6.4 Work Package Cost

The total work package cost represents the value spent to complete the work package. The higher

the total cost of the project, the bigger the work package size is. However, due to confidentiality

of such information the contractor did not include the costs in the given spreadsheets.

38
3.4.6.5 Resource Category

The resource category represents the type of labor resource that are utilized in the work package.

Each category includes all the different subcategories of the specified resource. For example,

construction estimator, construction inspector and senior project manager are different labor types,

but they are all categorized in one type as project manager. The labor resource categories in the

case study include 12 types: Project Management, Overburden Operator, Carpenter, Welder, HSE,

Ironworker, Labor, Operator, Pipefitter, Project Execution, Quality, and Accounting.

Frequency of Different Resources in Collected Dataset

Accounting
Carpenter
HSE
Ironworker
Labor
Operator
Overburden Operator
Pipefitter
Project Execution
Project Management
Quality
Welder
0 10 20 30 40 50 60 70

Figure 10. Distribution of Resource Categories

3.4.6.6 Key Quantity

The key quantity refers to the amount of main material, service, etc. that represents the scope of a

work package. Based on the work package type the unit of measure could be various. For instance,

the key quantity of a concrete work package is measured in cubic meters. The key quantity for

39
piping ie measured in inch/diameter. As a result, comparing key quantities is only acceptable when

they have the same unit of measurements.

3.4.6 Deficiencies in the Collected Dataset

There are a few limitations that were encountered during the data collection process. These

limitations are mostly due to the confidentiality or unavailability of some information from the

contractor side. Therefore, some assumptions were made to have a target dataset to forward in this

research. These limitations and assumptions include the following:

(1) The actual duration of each work package is calculated by subtracting the first date included in

the spreadsheet from the latest date which is not the accurate duration of the work package. In

some cases, there are several days with zero working hours in the first or last days of a work

package which is not clear whether the working hours are missed on those days, or they should

not be included in the duration. Therefore, it is assumed that there were no works on those days,

and they are excluded from the work package duration.

(2) The actual cost and budget were not available due to confidential issues. As a result, the case

study was done without cost attribute. Considering cost as in input would potentially improve the

performance of the forecasting model. However, the total number of working hours in each

package is available in the dataset.

(3) In calculating the duration, the holidays and weekends were excluded from the duration of the

work packages.

(4) In general work packages which include recurrent tasks such as maintenance and safety, the

durations were significantly large (near to project’s duration); so, in order to avoid high error

40
values in the models, these types of work packages were limited into a few records which had the

acceptable durations.

(5) The complexity of the work packages was not defined in the spreadsheets; and as there is not

a standard method for measuring the complexity, the complexity was also excluded from input

variables of the forecasting models.

(6) The data that is used for case study is limited into the projects which their work packages are

mostly limited into some specific types such as piping. Accordingly, this limitation does not allow

to train the model with different work packages of projects with different characteristics.

41
Chapter 4 Development of Forecasting Model for Labor Resources

4.1 Introduction

As previously mentioned, the objective of this study is to identify significant factors impacting

labor resource requirements using feature selection techniques, to develop a machine learning

model to predict the labor resources required for construction work packages to help project

managers in workforce allocation during the early stages of planning.

According to the applied methodology in the previous chapter, 6 factors affecting the labor

resource requirements were identified. In this chapter, a framework is proposed for development

of a forecasting model for labor resources where the identified factors are used as inputs.

To fulfil the objective of the research, an accurate and reliable forecasting models needs to be

developed by utilizing machine learning algorithms. Figure 11 indicates the steps of the following

framework to build the forecasting model. At the first step, the collected dataset is cleaned and

processed. Then, the clustering analysis is implemented to detect the optimum number of clusters

and their ranges. Also, the anomalies are identified in the dataset through outlier detection

techniques. Then, the feature selection methods are applied to investigate the feature with the huge

impacts on the performance of the prediction model. At the final step, different machine learning

algorithms are explored through their capabilities to select the most suitable algorithm for the

purpose of this study. The chosen algorithm is utilized in the development of the forecasting model.

Then, performance of the developed prediction model is evaluated by measuring the error value.

In this chapter, first data preprocessing and cluster analysis are explained. Then, feature selection

approach is elaborated. After that the selected machine learning algorithm and the developed

42
forecasting model are discussed. And finally, the performance evaluation of the forecasting model

and validation process are reviewed in detail.

Data Pre-processing Cluster Analysis Outlier Detection

Machine Learning Performance

Feature Selection Evaluation
Algorithm

Figure 11. Development of the Forecasting Model

4.2 Data Preprocessing

Most of the machine learning algorithms require some data cleaning and preprocessing prior to

any training. The performance of algorithms highly depends on the input data. So, the data needs

to be prepared correctly compatible with the algorithm to prevent any misleading or inaccurate

results One of the steps in data preparation is dealing with the missing values found in the dataset.

The missing information is mainly caused by improper tracking during the project. Also, outliers

can be found in the stored datasets where have been deficiencies in the tracking and storing

process. At the next step, the attributes in the dataset need to be scaled to prevent the forecasting

model from assigning unacceptable weights to the attributes because of greater values. Moreover,

the categorical attributes be set up. Handling the categorical attributes is another critical and

necessary part of preprocessing since some of the algorithms only take numerical inputs such as

regression and neural networks. Accordingly, the required modifications should be applied on the

dataset before proceeding into the training process.

43
So, proper formatting is necessary for the input dataset due to the discussed reasons. This process

includes some essential steps, in order to achieve legitimate input data for the model, which are

removing the records with missing values, eliminating outliers, handling categorical project

attributes, and normalizing the training dataset.

4.2.1. Missing Values

In the collected data some records did not include the hours spent on several days or the start and

end dates were not clear for some work packages. This occurred due to lack of allocation of hours

to projects on timesheets or the absence of timesheets to track the spent hours on some days.

Accordingly, some of the work packages were removed from the input data as it was unfeasible to

collect this information in any other way.

4.2.2. Cluster Analysis and Outliers Detection

As the duration of the work packages varies from a few days to several weeks, so it can obviously

be concluded that the model would not perform accurate prediction due to this huge variety. Hence,

a cluster analysis can improve the performance considerably. Table 1 indicates the error value for

the model before applying cluster analysis. As shown in the table, the Mean Absolute Error (MAE)

of the model is considerably high. Therefore, clustering and grouping the relevant work packages

is necessary for the performance enhancement. The details of the MAE and its calculations are

discussed later in Section 4.5.

Table 1. RNN Model Error Result Before Clustering

Number of MAE
Cluster Duration (day)
Work Packages (labor-hours)

1 223 >2 ± 14209

44
Cluster analysis is a statistical method for processing data. It works by organizing items into

groups, or clusters, on the basis of how closely associated they are. Cluster analysis is an

unsupervised learning algorithm, meaning that you do not know how many clusters exist in the

data before running the model. Unlike many other statistical methods, cluster analysis is typically

used when there is no assumption made about the likely relationships within the data. It provides

information about where associations and patterns in data exist, but not what those might be or

what they mean. Accordingly, it is a powerful data mining tool for identifying discrete groups of

work packages based on their durations.

k-means is commonly used algorithm in cluster analysis. k-means algorithm establishes the

presence of clusters by finding their centroid points. A centroid point is the average of all the data

points in the cluster. By iteratively assessing the Euclidean distance (Equation 1) between each

point in the dataset, each one can be assigned to a cluster. The centroid points are random to begin

with and will change each time as the process is carried out. Figure 12 indicates the process of

cluster analysis and outlier detection in the RapidMiner software. The result suggested that the

optimum number of clusters is 4. The boundary of each cluster is shown in the Table 2. It is

noteworthy that the outlier detection process is simultaneously applied with the clustering process.

The outliers are automatically removed from the data set.

𝑑(𝑝. 𝑞) = √∑𝑛𝑖=1(𝑞𝑖 − 𝑝𝑖 )2 𝑤ℎ𝑒𝑟𝑒 𝑝. 𝑞 = 𝑡𝑤𝑜 𝑝𝑜𝑖𝑛𝑡 𝑖𝑛 𝐸𝑢𝑐𝑙𝑖𝑑𝑒𝑎𝑛 𝑛 − 𝑠𝑝𝑎𝑐𝑒 (1)

45
 Figure 12. Clustering Analysis Process in RapidMiner

Table 2. Result of Cluster Analysis

Duration (day)
Cluster No. Number of Work Packages
Lower Bound Upper Bound

1 7 26 -

2 37 16 25

3 34 9 15

4 145 0 8

The following is a brief description of how clustering is done:

46
 1. First, the data is prepared for the analysis. Normalizing of the data is a typical step when

comparing attributes of different characteristics. In the proposed approach z-transformation

method is used to ensure that deviations are equal, so that the outlier has a clear meaning.

2. The clustering operator is applied to the data to find coherent groups in the dataset. Then,

the outliers are explored using the local outlier factor mechanism which identifies outliers

based on local outlier factors where locality is given by the k nearest neighbors. By

comparing the local density of a record and the density of its neighbors, the record is

considered as an outlier if its density is lower than the density of its neighbors.

3. The data is de-normalized by applying the reverse normalization model, thereby obtaining

the original dataset. Then, the examples are filtered to get one dataset with the outliers and

another with the rest.

4.2.3. Nominal Features

Dealing with categorical attributes might not be as easy as numerical ones. So, there are some

techniques to transform them into numerical variables. This transforming is an essential process

of data preparation. The prepared data set included two types of categorical data: nominal

attributes, such as work package type and resource category, and ordinal attribute like complexity.

Nominal ones consist of discrete categorical values that are labeled without order. On the other

hand, ordinal attributes include categorical values that have an order between them where the

distance between these values have an impact on the forecasting model.

One of the techniques to deal with these variables is to use dummy variables. Hence, dummy

coding is used to transform nominal attributes to numerical values. This method transforms the

categorical features into dummy variables. The dummy variable takes a value of 0 or 1. For

instance, new columns will be created for each work package type and the type will be specified

47
with a value of 1; the remaining columns will be 0 (Figure 13). All categories within each

categorical attribute are assigned a dummy value. This technique has an advantage over

transforming the values into unique integers, as the numerical values cannot be misinterpreted

from the prediction or the feature selection algorithms. However, the main disadvantage is that

more columns are added to the dataset (Elkholosy 2020).

Figure 13. Sample Illustration of Using Dummy Variables

Moreover, the work package complexity involved ordinal attributes with values of very low, low,

medium, high, very high. Therefore, assigning the proper numerical values to the 5-rating scale is

suitable for a good forecasting model. A value from 1 to 5 is assigned to the different levels of the

complexity, so the algorithms could recognize work packages with higher complexity should have

more weight allocated to them.

4.2.4. Normalizing & Data Splitting

After transforming categorical features into numerical values, all the values in the dataset became

numerical. The features varied in magnitude as each feature has a different unit. It is important to

scale features before training a machine learning model. Normalization is a common way of doing

this scaling. Subtract the mean and divide by the standard deviation of each feature.

48
The machine learning algorithms do not consider the units; they rely only the magnitude of the

values. Thus, all project features were standardized to allow the machine learning algorithm to

work properly. This is achieved by scaling each of project features to have a mean of zero and a

standard deviation of one. The mean and standard deviation should only be computed using the

training data so that the models have no access to the values in the validation and test sets. It is

also arguable that the model should not have access to future values in the training set when

training, and that this normalization should be done using moving averages.

Then, to provide an unbiased estimate of learning performance, the training data set is randomly

divided into 3 different subsets: (1) Training set is a subset of the dataset used to build predictive

models. In the developed model training subset consists of 70% of the records. (2) Validating set

is a subset of the dataset used to assess the performance of the model built in the training phase. It

provides a test platform for fine-tuning a model’s parameters and selecting the best performing

model. Validation set includes 20% of data set in the current model. (3) Testing set, or unseen data,

is a subset of the dataset used to assess the likely future performance of a model. If a model fits to

the training set much better than it fits the test set, overfitting is probably the cause. In the current

model 10% of data set is dedicated to the test subset.

The k-fold (k=8) cross validation technique is adopted for training and validation of the models.

As a result, the data records are randomly divided into 8 equal clusters and the train-validation

process is repeated for 8 runs, where each cluster is used for validation and the remaining clusters

are used to train the models. This process is repeated until all clusters are selected as the validation

set. This ensures that the dataset is completely utilized as the training set and leads to improvement

in the predictive performance of the model (Poh et al. 2018).

49
4.3 Feature Selection

Feature selection techniques are frequently used for data pre-processing purposes to enhance the

performance of the developed model through identifying the most relevant input variables (Poh et

al. 2018). In other cases, feature selection can reduce the computational time of the forecasting

model, such as when dealing with large dimensional datasets (Elkholosy 2020).

Two different algorithms are used to identify which attributes have a significant impact on the

labor resource requirements. First, a deterministic algorithm is applied, which behaves predictably.

The algorithm produces the same output every time, and the underlying machine passes through

the same sequence of states. Then, a greedy algorithm is applied. This algorithm follows the

problem-solving heuristic of making the locally optimal choice at each stage to find a global

optimum. The proposed feature selection approach includes two deterministic greedy algorithms

- ‘forward selection’ and ‘backward elimination’ - while some enhancements are added to the

standard algorithms which are described in the following.

With an initial population with n attributes in the input dataset. Each attribute is dedicated into

only one of the features. The attributes are evaluated by forward selection algorithm and only k

number of best ones are selected. For each of the k attribute sets the algorithm would do: if there

are j unused attributes, makes j copies of the attribute set and adds one of the previously unused

attributes to the attribute set. This step repeats if the performance is improved.

Backward elimination is the opposite of the forward selection. The algorithm starts with an

attribute set which uses all features. Then, evaluates all attribute sets and selects the best k. For

each of the k attribute sets the algorithm would do: if there are j attributes used, makes j copies of

50
the attribute sets, and removes one of the previously unused attributes from the attribute set. The

loop repeats as long as the performance is improved.

The method utilizes a machine learning algorithm to identify the best features trying different

combinations. After applying the forward selection and backward elimination, random forest

feature selection, which is a type of the embedded method, was used for feature selection. This

method trains a machine learning algorithm and then derives the features that impact the

prediction.

Prior to the feature selection process and transforming categorical data into numerical values, the

dataset included 6 attributes. As previously mentioned, the dummy coding process increased the

number of columns in the dataset. Thus, before implementing the feature selection process on the

dataset, the number of attributes increased to 27. Feature selection reduces the number of attributes

by eliminating irrelevant attributes and improves the performance of the forecasting models.

Feature selection is applied to each of the clusters that were developed using the k-means

algorithm. The selected features are evaluated using machine learning algorithms to measure their

performance and determine which features provide better results. For the defined problem, the

wrapper method identified the best features which provided the least errors in training and testing

the prediction models. Forward selection and backward elimination are used in the wrapper

method.

It is noteworthy that due to the limited attributes of the historical data, the feature selection

approach could not be implemented. The wrapper method requires a bigger dataset with more input

variables. Except the first cluster, other clusters did not have sufficient records for proper applying

51
of feature selection. Also, the first cluster had very limited input variables which made the feature

selection inapplicable.

4.4 Forecasting Models

Different prediction algorithms are explored to forecast the labor hours required for a work

package. Linear regression and Recurrent Neural Network (RNN) are the investigated algorithms

for the defined problem. Linear model is used due to its simplicity and ability to predict with

minimal error. RNN is one of the most powerful machine learning algorithms that are widely used

in different industries. Also, it is very formidable in dealing with complex problems and nonlinear

data.

The proposed model is trained by the collected historical data to identify inherent trends in the data

and unseen relationships between the output and input features. The model was initially trained to

forecast the required labor hours for a work package, using all work packages in the collected

dataset. But, the model did not perform acceptable and the error value was high. This huge error

was caused due to the wide range of work packages and their durations. Consequently, cluster

analysis was carried out as the possible solution to group work packages that have similar

durations. Finally, forecasting model was applied for each cluster, and its performance was

compared to the initial trial that included all work packages.

By implementing feature selection process, the significant impacting factors were identified. These

attributes are considered as the inputs for the forecasting models. They were used to train and test

the forecasting models to predict the desired output. Baseline model, linear regression, and RNN

models were investigated, and the performance was compared to determine which machine

learning algorithm fits best for the defined problem.

52
4.4.1. Baseline Model

Before building a trainable model, it would be good to have a performance baseline as a point for

comparison with the later more complicated models. This first task is to predict labor hour one

time unit in the future given the current value of all features. The current values include the current

labor hour. So, start with a model that just returns the current labor hour as the prediction,

predicting "No change". In other words, the baseline model's prediction is simply giving the value

of the previous time unit as the forecasted value for the next time unit. Obviously, the baseline

model will work less well for making a prediction further in the future.

Figure 14. Baseline Model Schema

4.4.2. Linear Model

The simplest trainable model is to insert linear transformation between the input and output. A

linear regression algorithm is developed for the model, and the performances is evaluated. The

process of building linear regression models does not require many configurations in the

hyperparameters compared to building a neural network model. The hyperparameters are the

model’s parameters whose values are used to control the models learning process. The model

requires numerical input variables, which are used to predict the labor hours by assigning weights

53
to every variable. The linear model does prediction of each time step independently from other

time units which means that each time step has its own linear model.

Figure 15. Linear Model Schema

4.4.3. Recurrent Neural Network Model

A single-time-step model like linear model has no context for the current values of its inputs. It

cannot see how the input features are changing over time. To address this issue the model needs

access to multiple time steps when making predictions (Figure 16).

54
 Figure 16. RNN Model Schema

Recurrent neural networks (RNN) are a class of neural networks that is powerful for modeling

sequence data such as time series or natural language. What makes RNN so powerful is the fact

that it does not take into consideration just the actual input but also the previous input which allows

it to memorize what happens previously. They are distinguished by their “memory” as they take

information from prior inputs to influence the current input and output. While traditional deep

neural networks assume that inputs and outputs are independent of each other, the output of

recurrent neural networks depend on the prior elements within the sequence. RNN unlike classic

algorithms such as naïve Bayesian, works well on sequence data because it takes the time step i as

input and combine with the output of time step i-1, the same thing would be applied for time step

i+1 and this is the reason it is called recurrent neural network because clearly the neural network

applies the same operations on each time step i of the sentence. The connections between the input

55
variables, hidden neurons, and output variables are composed of weights. These weights can vary

between negative and positive values. The importance of the input variable can be demonstrated

by values of the weights between that node and the hidden neurons.

Figure 17. RNN Model Structure

Another distinguishing characteristic of recurrent networks is that they share parameters across

each layer of the network. While feedforward networks have different weights across each node,

recurrent neural networks share the same weight parameter within each layer of the network. That

said, these weights are still adjusted in the through the processes of backpropagation and gradient

descent to facilitate reinforcement learning.

As previously mentioned, the RNN does predictions step-by-step. This means that in our model,

the algorithm is initially trained to forecast the required labor hours for the first day. Then the

56
network performs some enhancements automatically and is prepared for predicting the next day.

This cycle takes place until the algorithm reaches the last day within the work package.

4.5 Performance Evaluation

Basically, the estimated performance of a model tells how well it performs on unseen data.

Accordingly, after developing the forecasting model, model validation is required to ensure that

the model works properly on unseen data. As mentioned before, there are two approaches that are

used to evaluate the performance of the developed model. The first method involves splitting the

dataset into training set, validation set and a testing set; with the split percentage of 70%, 20% and

10% for training, validating, and testing, respectively. The other technique is k-folds cross

validation. This method splits the input data into k folds (k=8). The model is trained using the k–

1 fold and tested with the remained fold. This process takes place until all the records in the dataset

are included in the test set.

Besides, there are two common metrics used for evaluating the regression models namely, Root

Mean Squared Error and Mean Absolute Error (Narula and Wellington 1977). The Mean Absolute

Error (MAE) is the sum of the absolute differences between predictions and actual values

(Equation 2). On the other hand, Root Mean Squared Error (RMSE) measures the average

magnitude of the error by taking the square root of the average of squared differences between

prediction and actual observation (Equation 3). The MAE is the main criterion used to measure

the performance of the developed forecasting model.

1
𝑀𝐴𝐸 = × ∑𝑛1|𝑦𝑝𝑟𝑒𝑑 − 𝑦𝑟𝑒𝑓 | 𝑤ℎ𝑒𝑟𝑒 𝑦𝑝𝑟𝑒𝑑 = 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑣𝑎𝑙𝑢𝑒. 𝑦𝑟𝑒𝑓 = 𝑎𝑐𝑡𝑢𝑎𝑙 𝑣𝑎𝑙𝑢𝑒 (2)
𝑛

(𝑦𝑝𝑟𝑒𝑑 −𝑦𝑟𝑒𝑓 )2
𝑅𝑀𝑆𝐸 = √∑𝑛1 𝑛
𝑤ℎ𝑒𝑟𝑒 𝑦𝑝𝑟𝑒𝑑 = 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑣𝑎𝑙𝑢𝑒. 𝑦𝑟𝑒𝑓 = 𝑎𝑐𝑡𝑢𝑎𝑙 𝑣𝑎𝑙𝑢𝑒 (3)

57
Figure 18 indicates the MAE for the RNN model used in each cluster to forecast the labor hours.

The forecasting model for the first cluster performed with lower error in comparison to the other

clusters because the number of work packages used to train the models were greater than the

number of projects in the other clusters.

MAE of the Developed Forecasting Model

2500 160
140

No. of Work Packages

MAE (labor-hours)

2000
120

1500 100
80
1000 60
40
500
20
0 0
< 9 days 9 - 15 days 16 - 25 days > 25 days
Duration

RNN No. of Work Packages

Figure 18. MAE of the Forecasting Model

The performance of the forecasting models was considerably enhanced after omitting outliers,

applying cluster analysis, and selecting features prior to the development of the forecasting model.

Proposing one single forecasting model for all work packages was not acceptable and in conflict

with the initial problem definition of investigating the various durations and labor hours. With the

high error value and deficient performance, the result was not reliable for an estimate from the

58
model. Thus, a forecasting model was used for each cluster resulting in a noticeable decrease in

the error value.

After developing separate models, the errors of the models were still high, however they performed

much better than the baseline model. This concludes that the proposed models can perform well in

predicting the required labor hours as it is better than using the baseline model and estimates done

by the industry practitioners. Also, the models’ errors are expected to be reduced when the data is

collected in the proper manner as proposed in this research. Furthermore, the validation process

involved assessing model results by project managers and determining if the estimates provided

were reasonable. Also, their feedback was considered to improve the model’s performance and to

identify any drawbacks found in the model.

4.6 Summary

In this chapter, the taken steps of developing the model from data collection to performance

evaluation were discussed. As mentioned, data preparation was challenging as the data included

too much coded information from different sources. The collected data had much missing values

compared to the designed data acquisition model. Hence, the study was limited in different aspects.

Then, outlier detection methods were applied to eliminate the outliers found in the data set. The

cluster analysis also helped in detecting the anomalies.

Considering the wide variety of the work packages, developing one single model that can forecast

any construction work package was not applicable. Therefore, the clustering technique (k-means)

was applied to group the relevant work packages based on the duration. Then, a forecasting model

was trained for every cluster.

59
Two different robust machine learning algorithms (linear regression and RNN) were evaluated to

determine which algorithm is the best possible choice for the purpose of this study. Finally, the

RNN algorithm was chosen due to its high capability in forecasting time series and dealing with

anomalies. The MAE of the model was nearly high due to the limited data for training, but the

performance would be still acceptable and can be easily improved by adding more training data.

The error value for the first cluster which includes the work packages less than 9 days, has the

least error among the other clusters. The lower error is due to the higher number of training records.

Although, the overall mean absolute error of the model is nearly 4,900 labor-hours, but it is still

acceptable in comparison to the current industry practices. This is caused by the limited number

and diversity of work packages used for training and the lack of available information as input

variables. The accuracy of the models is expected to increase when the data is collected based on

the proposed data acquisition model, and by increasing the number of work packages to train the

RNN model.

60
Chapter 5 Implementation of the Labor Resource Forecasting Model

5.1 Introduction

The construction companies mostly do not take the advantages of their historical data. Besides,

professional complex software can be complicated and confusing. Having a simple forecasting

application can easily tackle these issues.

During this research, a simple application for estimating the labor hours for construction work

packages is developed. The application utilizes the developed forecasting model and provides the

prediction based on the user’s inputs. Figure 19 shows the interface of the application. The

interface is designed in a way to provide a simple and straight forward service into the user whether

it is a project manager or a researcher. Also, the output is presented decently so that it can be easily

interpreted and utilized. Developing such computer application can bring many advantages

including: (1) easy to run and no need for complex installation procedure. (2) providing a user-

friendly interface for utilizing the developed model. (3) Presenting a framework to aid project

analysis and realistic future planning. (4) Avoiding users interfere with the back-end codes. (5)

Providing a decent visualization for the output of the model.

61
Figure 19. Interface of the Application

62
Figure 20. Sample Run of the Application

63
5.2 Application Components

The development of the application was an updating process. Figure 21 presents the initial version

of the interface. The feedback from experienced project managers provided useful modifications

to make the application useable for industry experts. The initial idea was to develop a generic

application to support different work packages and resources. But limitations in time and data

prevented developing such generic application and the current version is limited to the collected

data from one contractor.

Figure 21. Initial Version of the Interface

The application includes two main parts. Following elaborates these two major components of the

current application:

64
1. The main window and dropdown buttons (Figure 22):

To predict the labor resource requirements, the user selects the required inputs. After getting

user’s inputs, the application runs the trained RNN model.

Figure 22. Main Window and Dropdown Buttons (Inputs)

There are 3 dropdown buttons for the work package category, resource, and duration. Figures

23, 24 and 25 represent these dropdowns and their setup codes.

65
 Figure 23. Dropdown Buttons and Lists

Figure 24. Main Frame Setup Code

It is noteworthy that the categories of work packages and resources are limited to the construction

ones and the other types such as management, planning and procurement are deleted from the lists.

This is due to the different characteristic of these work packages. They usually last much longer

in comparison to the construction work packages and has consistent labor hours spent each day.

For example, a project management work package would last 200 days with a steady labor hour

for each day while an excavation work package lasts mostly less than 30 days with a fluctuating

required labor hours during the execution. As a result, to achieve a more accurate and beneficial

model, the general work packages and resources are eliminated from the model and only

construction ones are included in the proposed application.

66
 Figure 25. Dropdown Buttons Setup Code

2. The output and plotting:

As previously discussed, a forecasting model is developed for different work packages

clustered based on their durations. One of the trained recurrent neural networks is used to

forecast the labor resource requirement of a new work package, and the selection of the cluster

67
is made when the user selects the duration. To avoid confusion, the codes are not accessible

within application and only developers have access to modify the models and their parameters.

Figures 26 and 27 represent the sample codes for the output setup.

Figure 26. Model's Output Setup Code

Figure 27. Output Plotting Setup Code

The model provides the user with various outputs including the graph of labor hours predicted

for each working day by the RNN model, the total labor hours estimated for the work package,

and the error value of the predicted hours which can be interpreted as a confidence interval.

The output would provide a precise perspective about the labor hours needed for a given work

package. For instance, Figure 28 represents the required duration and labor hours for a short

68
 steel work package. According to the result, the work package needs 4 days for completion

and a total of 162.5 labor hours with the MAE value equal to 6 labor hours.

Figure 28. Model's Prediction for Required Labor Resource in a Steel Work Package

In another example, Figure 29 shows the predicted labor hours for a piping work package with

the duration of 17 days while Figure 30 elaborates the real labor hours spent for a piping work

package with similar characteristics.

69
Figure 29. Predicted Labor Hours Required for a Piping Work Package

70
 40
35
30

Labor-Hours
25
20
15
10
5
0
0 5 10 15 20
Time (Day)

Figure 30. Actual Labor-Hours Spent for a Piping Work Package

As shown in Figure 29, the required hours for a piping work package with the duration of 17 days

is predicted as 260 labor-hours. On the other hand, Figure 30 shows the spent labor hours for a

similar piping work package which lasted 18 days. The actual total hours is 288 labor-hours which

is near to the total o predicted hours. This comparison indicates that the developed application

works properly, and its accuracy is acceptable. But it is clear that the two curves are not similar

and in some of the time steps the difference between the actual and predicted hours is high. The

reason of this error can be found in the inconsistency of the training dataset and the high variance

in the actual labor hours spent in each time step which can lead to the confusion of the model and

accordingly inaccurate estimation of the required labor hours for each day. By implementing the

training process of the application with the proposed data acquisition model, such inconsistencies

are reduced, and the performance is expected to improve.

71
Chapter 6 Summary, Limitations and Future Work

6.1 Research Summary

Considering the high rates of labor resources in construction projects, proposing a decent labor

resource management framework is critical. Predicting needed labor requirements can be a

substantial step towards labor resource management. With the recent developments in the area of

artificial intelligence and machine learning, these technologies can potentially be adopted to

develop resource prediction models. As the labor hours are often documented and recorded in

historical datasets, there is a potential to use these datasets to analyze labor resources, identify

underlying patterns, and forecast future trends.

Accordingly, this research aims to combine the benefits of artificial intelligence and historical data

of previous projects to identify the significant factors affecting the labor hour requirements and to

develop a generic forecasting model. This objective is achieved by using machine learning

algorithms and developing neural network models for various construction work packages.

Although, the error of the models is high due to the wide range of work packages and the limited

number of records used to train the models.

In the first step, the key factors impacting the labor resource requirements were explored through

previous studies and industry practices. Therefore, an initial list of attributes was provided. Then,

historical data of previously performed construction projects was collected focusing on the

explored key attributes. The proposed list of features consisted of the following attributes: work

package type, resource category, duration, cost, complexity, and key quantity.

The process of data collection was challenging since some of the target attributes were not

available by the contractors. The data was extracted from the tracking spreadsheets used by project

72
managers. These pre-defined spreadsheets included the information of the tasks and activities

happening each day. But all the information was not accessible due to confidentiality. As a result,

the collected dataset was limited into a small portion of the original dataset.

In the next step, the collected data was pre-processed and cleaned to be used as the training set for

the models. Two machine learning algorithms (linear regression and RNN) were evaluated for

using in the forecasting models. It was concluded that the recurrent neural network algorithm is

more capable in predicting time series and was chosen as the final algorithm. Because of the wide

range of the work packages, developing one forecasting model for all of them was not applicable

and would lead to high error in estimated labor hours (Table 1). Hence, clustering analysis was

performed on the dataset to group work packages based on their durations (Figure 31). Then,

forecasting models were developed for each cluster. The comparison of the results (Table 1 and 3)

clearly indicated that clustering has reduced the error value and increased the accuracy of the

model. Also, a feature selection method was designed to identify the most impacting factors. But

due to the limited attributes of the collected dataset, the feature selection was not applied.

Clusters & Durations

> 25 days
3%

16 - 25 days
17% 0 - 8 days
9 - 15 days
9 - 15 days
15% 0 - 8 days 16 - 25 days
65%
> 25 days

Figure 31. Cluster Analysis Result

73
The performance of the RNN model was evaluated and the results is presented in Table 3. The

developed model for the first cluster – work packages with less than 9 days -had the lowest error

among the other clusters due to the efficient number of the training records. The mean absolute

error of the cluster is 209 labor-hours which is still high but can be improved by adding high-

quality training dataset. The high value for the error is mainly due to the limited number of

available records and the lack of some attributes as the models’ inputs. The accuracy would be

improved by increasing the training dataset records and collecting information according to the

provided data acquisition model.

Table 3. RNN Model Clusters and Results

Number of Duration MAE

Cluster
Work Packages (day) (labor-hours)

1 145 0-8 ± 209

2 34 9 - 15 ± 1841

3 37 16 - 25 ± 636

4 7 ≥ 26 ± 2245

The output of the developed models is the daily labor resource requirements estimated for a

specific type of work package. The result of the model is not a single value but the distribution of

the required labor hours over the work package duration. Consequently, the output would be more

reliable and project managers could benefit more from the result as it is not limited into only an

estimated total hour for the whole work package.

Eventually, a computer application was designed for the proposed model. The application was

developed to provide a user-friendly interface, so that the user – whether a senior project manager

74
or a researcher - can easily benefit from the developed forecasting model. By running the

application, the user can select the desired inputs and after running the trained model at the back

end, the predicted result is plotted with the estimated error value.

6.2 Limitations

While this research aims to develop a forecasting model for labor resource requirements, there

were some limitations which can be summarized as follows:

The most important during this research was the data collection process. Finding the appropriate

contractor was challenging as there are not many contractors who are willing to share their

historical data due to confidential issues and their policies. Also, the pandemic and lockdown

situation highly impacted the communications and reaching out to industry partners became

difficult. This situation caused delay in the data collection process.

After finding the contractor only a limited dataset was received. The shared dataset was including

only few attributes of only one construction project. This small dataset did not allow this research

to fully benefit from the proposed approach. The proposed feature selection method could not be

applied since the input variables were limited. Moreover, the input variables were limited to 4

attributes as the other expected inputs were not available. All these issues lead to decreased

accuracy in the performance of the prediction models.

The quality of the training dataset was not of high quality. As mentioned, the dataset was limited

to the construction projects with similar characteristics including only specific types of work

packages. Also, some of their work packages were general ones which had long durations with

constant daily labor hours. Such work packages would disturb the training process and cause

inaccurate model.

75
The machine learning algorithms especially RNN, need very high number of records for training

in order to perform predictions with low error. Otherwise, they are not capable to forecast the labor

hours precisely and cannot be reliable. Consequently, the performance of the developed model is

impacted and needs much more high-quality historical data for training to improve its high error

predictions.

6.3 Future Work and Recommendations

This research can be expanded in different directions. The current model can be expanded to

forecast more work package types and to consider other resource categories. Collecting more

historical data from different construction projects and supplying high-quality dataset to the model

can improve the performance and reliability. Also, adding more significant factors as inputs to the

model can potentially increase the accuracy.

The developed model can be integrated into a greater model which is capable to forecast different

aspects of a work package such as duration, required labor resources, key quantity etc. With the

rapid advancement of artificial intelligence, new algorithms with more capabilities in forecasting

are emerging. RNN is just a sample of these brand-new algorithms which was used in this research.

The new algorithms and techniques are always potential options for improvement of the current

models.

76
References

Agapiou, A., Price, A. D., & McCaffer, R. (1995). Planning future construction skill requirements:
understanding labour resource issues. Construction Management and Economics, 13(2), 149-161.

Alizadeh, S. S., Mortazavi, S. B., & Mehdi Sepehri, M. (2015). Assessment of accident severity
in the construction industry using the Bayesian theorem. International journal of occupational
safety and ergonomics, 21(4), 551-557.

Amiri, M., Ardeshir, A., Fazel Zarandi, M. H., & Soltanaghaei, E. (2016). Pattern extraction for
high-risk accidents in the construction industry: a data-mining approach. International journal of
injury control and safety promotion, 23(3), 264-276.

Amrutha, V. N., & Geetha, S. N. (2020). A systematic review on green human resource
management: Implications for social sustainability. Journal of Cleaner Production, 247, 119131.

and Computer Engineering. IEEE. p. 1–6

Antonioli, D., Mancinelli, S., Mazzanti, M. (2013). Is environmental innovation embedded within
high-performance organisational changes? The role of human resource management and
complementarity in green business strategies. Res. Policy 42 (4), 975–988.

Apaydin, H., Feizi, H., Sattari, M. T., Colak, M. S., Shamshirband, S., & Chau, K. W. (2020).
Comparative analysis of recurrent neural network architectures for reservoir inflow forecasting.
Water, 12(5), 1500.

Arnett, D. B., Laverie, D. A., & McLane, C. (2002). Using job satisfaction and pride as internal-
marketing tools. Cornell hotel and restaurant administration quarterly, 43(2), 87-96.

Baiden, B.K., Price, A.D., 2011. The effect of integration on project delivery team effectiveness.
Int. J. Proj. Manag. 29 (2), 129–136.

Ballesteros Pérez, P., González-Cruz, Mª.C., Pastor-Ferrando, J.P. (2010). Analysis of

construction projects by means of value curves. Int. J. Proj. Manag. Volume 28 (Issue 7), 719–
731.

77
Baloh, P., Desouza, K.C., Hackney, R. (2012). Contextualizing organizational interventions of
knowledge management systems: a design science perspective. J. Am. Soc. Inf. Sci. Technol. 63
(5), 948–966. Partington, D., Pellegrinelli, S., Young, M., 2005. Attributes and levels of
programme management competence: an interpretive study. Int. J. Proj. Manag. 23 (2), 87–95.

Baron, R. A. (2003). Human resource management and entrepreneurship: Some reciprocal benefits
of closer links. Human Resource Management Review, 2(13), 253-256.

Becker, B., & Gerhart, B. (1996). The impact of human resource management on organizational
performance: Progress and prospects. Academy of management journal, 39(4), 779-801.

Becker, B.E., Huselid, M.A. (2006). Strategic human resources management: where do we go from
here? J. Manag. 32 (6), 898–925.

Belout, A., Gauvreau, C. (2004). Factors influencing project success: the impact of human resource
management. Int. J. Proj. Manag. 22 (1), 1–11.

Bonaccorso, G. (2018). Machine Learning Algorithms: Popular algorithms for data science and
machine learning. Packt Publishing Ltd.

Bourne, L., Walker, D.H. (2005). The paradox of project control. Team Perform. Manag. 11 (5–
6), 157–178.

Bowen, D.E., Ostroff, C. (2004). Understanding HRM-firm performance linkages: the role of the
"Strength" of the HRM system. Acad. Manag. Rev. 29 (2), 203–221.

Bredin, K. (2008). People capability of project-based organizations: a conceptual framework. Int.

J. Proj. Manag. 26 (5), 566–576.

Briscoe, G., & Wilson, R. A. (1993). Employment forecasting in the construction industry.
Ashgate Publishing.

Buller, P.F., McEvoy, G.M., 2012. Strategy, human resource management and performance:
sharpening line of sight. Hum. Resour. Manag. Rev. 22 (1), 43–56.

Bush, V. G. (1973). Construction management: A handbook for contractors, architects, and

students. Reston Publishing.

78
Campion, M.A., Medsker, G.J., Higgs, A.C. (1993). Relations between work group characteristics
and effectiveness: implications for designing effective work groups. Pers. Psychol. 46 (4), 823–
847

Chan, A.P.C., Chiang, Y.H., Mak, S.W.K., Choy, L.H.T. and Wong, J.M.W. (2006) Forecasting
the demand for construction skills in Hong Kong, Construction Innovation, 6(1), 3-19.

Chan, A.P.C., Wong, J.M.W. and Chiang, Y.H. (2003) Modelling labour demand at project level
– an empirical study in Hong Kong, Journal of Engineering, Design and Technology, 1(2), 135-
50.

Chen, L. H., Liaw, S. Y., & Lee, T. Z. (2003). Using an HRM pattern approach to examine the
productivity of manufacturing firms–an empirical study. International Journal of Manpower.

Chiang, Y. H., Wong, F. K. W., & Liang, S. (2018). Fatal construction accidents in Hong Kong.
Journal of construction engineering and management, 144(3), 04017121.

Chukwu, A. U., & Adepoju, K. A. (2012). On the power efficiency of artificial neural network
(ANN) and the classical regression model. Progress in Applied Mathematics, 3(2), 28-34.

Cleland, D.I., Ireland, L.R. (2006). Project Management: Strategic Design and Implementation,
5th ed. McGraw-Hill Professional, New York.

Dang, Y., Zhang, Y., Zhang, D., & Zhao, L. (2005, August). A KNN-based learning method for
biology species categorization. In International Conference on Natural Computation (pp. 956-
964). Springer, Berlin, Heidelberg.

Datta, D.K., Guthrie, J.P., Wright, P.M. (2005). Human resource management and labor
productivity: does industry matter? Acad. Manag. J. 48 (1), 135–145.

Davis, K. (2014). Different stakeholder groups and their perceptions of project success. Int. J. Proj.
Manag 32 (2), 189–201.

Dessler, G. (2012). A Framework for Human Resource Management, 7th ed. Prentice Hall,
Pearson.

79
Druker, J., White, G., Hegewisch, A., & Mayne, L. (1996). Between hard and soft HRM: human
resource management in the construction industry. Construction Management & Economics,
14(5), 405-416.

Duda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern Classification. John Wiley & Sons. Inc.,
New York, 2.

Ehrenberg, R. G., & Smith, R. S. (2003). Modern Labor Economics: theory and public policy eight
edition.

Franco-Lopez, H., Ek, A. R., & Bauer, M. E. (2001). Estimation and mapping of forest stand
density, volume, and cover type using the k-nearest neighbors method. Remote sensing of
Environment, 77(3), 251-274.

Ganesan, S., Hall, G., and Chiang, Y. H. (1996). Construction in Hong Kong: Issues in labor supply
and technology transfer, Aldershot, Avebury, U.K.

Garavan, T. N., & Morley, M. J. (2006). Re-dimensionalising boundaries in the theory and practice
of Human Resource Development. International Journal of Learning and Intellectual Capital, 3(1),
3-13.

Gerassis, S., Martín, J. E., García, J. T., Saavedra, A., & Taboada, J. (2017). Bayesian decision
tool for the analysis of occupational accidents in the construction of embankments. Journal of
construction engineering and management, 143(2), 04016093.

Gershenson, C. (2003). Artificial neural networks for beginners. arXiv preprint cs/0308031.

Gidado, K. I., & Millar, A. J. (1992, September). The effect of simple overlap of the stages of
building construction on the project complexity and contract time. In Proceedings of the 8 th
Annual Conference, Association of Researchers in Construction Management (pp. 307-317).

Gondia, A., Siam, A., El-Dakhakhni, W., & Nassar, A. H. (2020). Machine Learning Algorithms
for Construction Projects Delay Risk Prediction. Journal of Construction Engineering and
Management, 146(1), 04019085.

Gorunescu, F. (2011). Data Mining, San Francisco, CA, itd. Morgan Kaufmann. doi, 10, 978-3.

Gould, F. E. (2002). Construction project management, Prentice-Hall, Upper Saddle River, N.J.

80
Gruneberg, S.L. (1997), Construction Economics – An Introduction, Macmillan, London

Guo, G., Wang, H., Bell, D., Bi, Y., & Greer, K. (2003, November). KNN model-based approach
in classification. In OTM Confederated International Conferences" On the Move to Meaningful
Internet Systems" (pp. 986-996). Springer, Berlin, Heidelberg.

Gurmu, A. T., & Ongkowijoyo, C. S. (2020). Predicting Construction Labor Productivity Based
on Implementation Levels of Human Resource Management Practices. Journal of Construction
Engineering and Management, 146(3), 04019115.

Hammad, A. M. (2009). An integrated framework for managing labour resources data in industrial
construction projects: A Knowledge Discovery in Data (KDD) approach (Vol. 71, No. 01).

Han, J., Pei, J., & Kamber, M. (2011). Data mining: concepts and techniques. Elsevier.

Handy, C. B. (1985). Understanding organizations, 3rd Ed., Penguin, Harmondsworth, U.K

Fairweather, V. (1986). “Record high-rise, record low steel.” Civ. Eng. (N.Y.), 56(8), 42–45

Haykin, S. (1999). Learning processes; single-layer perceptrons; multilayer perceptrons. Neural

Networks A Comprehensive Foundation. 2nd ed. USA: Prentice Hall International Inc, 14-68.

Haykin, S. (2007). Neural networks: a comprehensive foundation. Prentice-Hall, Inc.

Hendry, D.F. (1985), Empirical Modelling in Dynamic Econometrics, Nuffield College, Oxford.

Heravi, G., & Eslamdoost, E. (2015). Applying artificial neural networks for measuring and
predicting construction-labor productivity. Journal of Construction Engineering and Management,
141(10), 04015032.

Hevner, A.R. (2007). The three-cycle view of design science research. Scand. J. Inf. Syst. 19 (2),
87.

Hibat-Allah, M., Ganahl, M., Hayward, L. E., Melko, R. G., & Carrasquilla, J. (2020). Recurrent
neural network wave functions. Physical Review Research, 2(2), 023358.

Huang, C. H., & Hsieh, S. H. (2020). Predicting BIM labor cost with random forest and simple
linear regression. Automation in Construction, 118, 103280.

81
Huemann, M. (2010). Considering human resource management when developing a project-
oriented company: case study of a telecommunication company. Int. J. Proj. Manag. 28 (4), 361–
369.

Huselid, M. A. (1995). The impact of human resource management practices on turnover,

productivity, and corporate financial performance. Academy of management journal, 38(3), 635-
672.

Inyang, B. J. (2011). Creating value through people: Best human resource (HR) practices in
Nigeria. International business and management, 2(1), 141-150.

Iwu, C. G. (2016). Effects of the use of electronic human resource management (E-HRM) within
human resource management (HRM) functions at universities. Acta Universitatis Danubius.
Administratio, 8(1).

Jantan, H., Hamdan, A. R., & Othman, Z. A. (2009). Knowledge discovery techniques for talent
forecasting in human resource application. World Academy of Science, Engineering and
Technology, 50, 775-783.

Katz, H. C., Kochan, T. A., Keefe, J. H., Lazear, E., & Eads, G. C. (1987). Industrial relations and
productivity in the US automobile industry. Brookings papers on economic activity, 1987(3), 685-
727.

Kubat, M., & Cooperson Jr, M. (2001). A reduction technique for nearest-neighbor classification:
Small groups of examples. Intelligent Data Analysis, 5(6), 463-476.

Kukenberger, M.R., Mathieu, J.E., Ruddy, T. (2012). A cross-level test of empowerment and
process influences on members' informal learning and team commitment. J. Manag.

Lall, U., & Sharma, A. (1996). A nearest neighbor bootstrap for resampling hydrologic time series.
Water resources research, 32(3), 679-693.

Laufer, A., Woodward, H., & Howell, G. A. (1999). Managing the decision-making process during
project planning. Journal of Management in Engineering, 15(2), 79-84.

82
Lee, B. H., & Scholz, M. (2006). A comparative study: Prediction of constructed treatment wetland
performance with k-nearest neighbors and neural networks. Water, Air, and Soil Pollution, 174(1-
4), 279-301.

Lemessany, J., & CLAPP, M. A. (1978). Resource inputs to construction: the labour requirements
of house building. Building Research Establishment.

Li, H. (1996). Case-based reasoning for intelligent support of construction negotiation.

Information & Management, 30(5), 231-238.

Lim, C.S., Mohamed, M.Z. (1999). Criteria of project success: an exploratory reexamination. Int.
J. Proj. Manag. 17 (4), 243–248.

Liu, M., & Ballard, G. (2008, July). Improving labor productivity through production control. In
Proceedings of the 11th Annual Conference of International Group for Lean Construction.

Loosemore, M., Dainty, A., & Lingard, H. (2003). Human resource management in construction
projects: strategic and operational approaches. Taylor & Francis.

Marchington, M., & Wilkinson, A. (2002). People management and development: Human resource
management at work. CIPD Publishing.

McClelland, J. L., Rumelhart, D. E., & PDP Research Group. (1986). Parallel distributed
processing. Explorations in the Microstructure of Cognition, 2, 216-271.

McConnell, C. R., Brue, S. L., and Macpherson, D. A. (2003). Contemporary labor economics,
McGraw-Hill, London.

Meharie, M. G., & Shaik, N. (2020). Predicting Highway Construction Costs: Comparison of the
Performance of Random Forest, Neural Network and Support Vector Machine Models. Journal of
Soft Computing in Civil Engineering, 4(2), 103-112.

Mourya, S. K., & Gupta, S. (2012). Data mining and data warehousing. Alpha Science
International, Ltd.

Muizz O. Sanni-Anibire, Rosli Mohamad Zin & Sunday Olusanya Olatunji (2020): Machine
learning model for delay risk assessment in tall building projects, International Journal of
Construction Management

83
Mutua, A. M. (2019). The Impact of Human Resource Planning On Organizational Performance
(Doctoral dissertation, University of Nairobi).

Nakanishi, Y. (2001). Dynamic labor demand using error correction model. Applied Economics,
33(6), 783-790.

Paliwal, M., & Kumar, U. A. (2009). Neural networks and statistical techniques: A review of
applications. Expert systems with applications, 36(1), 2-17.

Pant, I., Baroudi, B. (2008). Project management education: the human skills imperative. Int. J.
Proj. Manag. 26 (2), 124–128.

Patterson, D. (1996) Artificial Neural Networks. Prentice Hall, Singapore.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... & Duchesnay,
E. (2011). Scikit-learn: Machine learning in Python. the Journal of machine Learning research, 12,
2825-2830.

Persad, K. R., O'Connor, J. T., & Varghese, K. (1995). Forecastng Engineering Manpower
Requirements for Highway Preconstruction Activities. Journal of Management in Engineering,
11(3), 41-47.

Peterson, L. E. (2009). K-nearest neighbor. Scholarpedia, 4(2), 1883.

Pfeffer, J. (1998). The Human Equation: Building Profits by Putting People First. Harvard
Business Review Press, Boston.

Pinker, E. J., & Larson, R. C. (2003). Optimizing the use of contingent labor when demand is
uncertain. European Journal of Operational Research, 144(1), 39-55.

Pinto, J.K., Prescott, J.E. (1988). Variations in critical success factors over the stages in the project
life cycle. J. Manag. 14 (1), 5–18.

Poh, C. Q., Ubeynarayana, C. U., & Goh, Y. M. (2018). Safety leading indicators for construction
sites: A machine learning approach. Automation in construction, 93, 375-386.

Pournader, M., Tabassi, A. A., & Baloh, P. (2015). A three-step design science approach to
develop a novel human resource-planning framework in projects: the cases of construction projects
in USA, Europe, and Iran. International journal of project management, 33(2), 419-434.

84
Probst, P., Wright, M. N., & Boulesteix, A. L. (2019). Hyperparameters and tuning strategies for
random forest. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 9(3),
e1301.

Project Management Institute, (2013). A Guide to the Project Management Body of Knowledge
(PMBOK Guide), 5th ed. Project Management Institute, Newtown Square, Pennsylvania

Raiden, A.B., Dainty, A.R., Neale, R.H., (2004). Current barriers and possible solutions to
effective project team formation and deployment within a large construction organisation. Int. J.
Proj. Manag. 22 (4), 309–316.

Reilly, P. (2003). Guide to workforce planning in local authorities, employers’ organization for
local government, London.

Rosa, J. L. A., Ebecken, N. F., & Costa, M. C. A. (2003). Towards on an optimized parallel KNN-
fuzzy classification approach. WIT Transactions on Information and Communication
Technologies, 29.

Scott-Young, C., Samson, D. (2008). Project success and project team management: evidence from
capital projects in the process industries. J. Oper. Manag. 26 (6), 749–766.

Song, Y., Huang, J., Zhou, D., Zha, H., & Giles, C. L. (2007, September). Iknn: Informative k-
nearest neighbor pattern classification. In European Conference on Principles of Data Mining and
Knowledge Discovery (pp. 248-264). Springer, Berlin, Heidelberg.

Tabassi, A.A., Ramli, M., Bakar, A.H.A., (2012). Effects of training and motivation practices on
teamwork improvement and task efficiency: the case of construction firms. Int. J. Proj. Manag. 30
(2), 213–224

Tam, V. W., Tam, C. M., Chan, J. K., & Ng, W. C. (2006). Cutting construction wastes by
prefabrication. International Journal of Construction Management, 6(1), 15-25.

Tang, J.C.S., Karasudhi, P. and Tachopiyahoon, P. (1990), “Thai construction industry: demand
and projection”, Construction Management and Economics, Vol. 8, pp. 249-57.

Tixier, A. J. P., Hallowell, M. R., Rajagopalan, B., & Bowman, D. (2016). Application of machine
learning to construction injury prediction. Automation in construction, 69, 102-114.

85
Tommelein, I. D., Levitt, R. E., & Hayes-Roth, B. (1992). Site-layout modeling: how can artificial
intelligence help? Journal of construction engineering and management, 118(3), 594-611.

Tsui, A.S. (1987). Defining the activities and effectiveness of the human resource department: a
multiple constituency approach. Hum. Resour. Manag. 26 (1), 35–69

Tsui, A.S., Milkovich, G.T. (1987). Personnel department activities: constituency perspectives and
preferences. Pers. Psychol. 40 (3), 519–537

Turner, J. R., & Müller, R. (2005). The project manager's leadership style as a success factor on
projects: A literature review. Project management journal, 36(2), 49-61.

Turner, P. (2002). Strategic human resource forecasting. London: CIPD.

Wang, M., Cui, Y., Wang, X., Xiao, S., & Jiang, J. (2017). Machine learning for networking:
Workflow, advances and opportunities. Ieee Network, 32(2), 92-99.

Wang, Y., & Yao, Y. (1999). Sources of China’s economic growth, 1952–99: incorporating human
capital accumulation. The World Bank.

Wauters, M., & Vanhoucke, M. (2017). A nearest neighbour extension to project duration
forecasting with artificial intelligence. European Journal of Operational Research, 259(3), 1097-
1111.

Werther, W. B., & Davis, K. (1982). Personnel management and human resources. New York:
McGraw-Hill.

Widyanty, W., Daito, A., Riyanto, S., & Nusraningrum, D. (2020). Gaining a competitive
advantage through strategic human resource management in Indonesian construction industry.
Management Science Letters, 10(9), 2021-2028.

Wild, A. (2002). The unmanageability of construction and the theoretical psycho‐social dynamics
of projects. Engineering Construction and Architectural Management, 9(4), 345-351.

Witten IH, Frank E, Hall MA. (2011). Data mining: practical machine learning tools and
techniques. Burlington, Massachusetts, USA: Morgan Kaufmann Publishers.

Wong, J., Chan, A., & Chiang, Y. H. (2004). A critical review of forecasting models to predict
manpower demand. Construction Economics and Building, 4(2), 43-56.

86
Wong, J.M.W., Chan, A.P.C. and Chiang, Y.H. (2003), “Determinants of construction manpower
demand: a review from literature and practitioners’ experience”, 2nd International Conference on
Construction in the 21st Century (CITC-II), Hong Kong, 10-12 December 2003, pp. 158-63.

Wright, P.M., Boswell, W.R. (2002). Desegregating HRM: a review and synthesis of micro and
macro human resource management research. J. Manag. 28 (3), 247–276

Wu, J. D., & Chan, J. J. (2009). Faulted gear identification of a rotating machinery based on
wavelet transform and artificial neural network. Expert Systems with Applications, 36(5), 8862-
8875.

Xie, Q. (2020). Machine learning in human resource system of intelligent manufacturing industry.
Enterprise Information Systems, 1-21.

Yaseen, Z. M., Ali, Z. H., Salih, S. Q., & Al-Ansari, N. (2020). Prediction of risk delay in
construction projects using a hybrid artificial intelligence model. Sustainability, 12(4), 1514.

Zaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv
preprint arXiv:1409.2329.

Zhou, X., Zhu, X., Dong, Z., & Guo, W. (2016). Estimation of biomass in wheat using random
forest regression algorithm and remote sensing data. The Crop Journal, 4(3), 212-219.

Zhu, X., Zhang, S., Jin, Z., Zhang, Z., & Xu, Z. (2010). Missing value estimation for mixed-
attribute data sets. IEEE Transactions on Knowledge and Data Engineering, 23(1), 110-121.

Zwikael, O., Unger-Aviram, E. (2010). HRM in project groups: the effect of project duration on
team development effectiveness. Int. J. Proj. Manag. 28 (5), 413–421.

87
Appendix

The following figures elaborate some parts of the python code for developing the application:

Figure A. Loading the Libraries

Figure B. Setting up the Main Window of the Interface (size, title, etc.)

88
 Figure C. Defining Dropdown Buttons and their Values

Figure D. Defining the Labels and Texts - Placing the Buttons and their Labels

89
 Figure E. Defining the Different Categories of Duration Based on the Result of Clustering

Figure F. Filtering the Records Based on the Selected Attributes by User

Figure G. Dropping Additional Attributes and Filtering Records Based on Selected Duration

Figure H. Calculating the Duration

Figure I. k-fold Cross Validation for Training and Testing – Calculating Error Value

90
Figure J. Setting up the Canvas for the Predicted “Total Hour” and “Error”

Figure K. Plot Configuration and Setting up a Canvas to Elaborate the Plot

91