Building Research & Information

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/rbri20

A construction cost estimation framework using

DNN and validation unit

Salman Saeidlou & Nikdokht Ghadiminia

To cite this article: Salman Saeidlou & Nikdokht Ghadiminia (2023): A construction cost
estimation framework using DNN and validation unit, Building Research & Information, DOI:
10.1080/09613218.2023.2196388

To link to this article: https://doi.org/10.1080/09613218.2023.2196388

© 2023 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group

Published online: 11 Apr 2023.

Submit your article to this journal

Article views: 664

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=rbri20
 BUILDING RESEARCH & INFORMATION
https://doi.org/10.1080/09613218.2023.2196388

A construction cost estimation framework using DNN and validation unit

Salman Saeidloua and Nikdokht Ghadiminiab
a
School of Engineering, Technology and Design, Canterbury Christ Church University, Canterbury, United Kingdom; bSchool of Architecture,
Computing, and Engineering, University of East London, London, United Kingdom

ABSTRACT ARTICLE HISTORY

Accurate construction cost estimation is crucial to completing projects within the planned Received 14 November 2022
timeframe and expenditure. The estimation process depends on multiple variables maintaining Accepted 9 March 2023
complex relationships between themselves and the target cost. As a result, an in-depth analysis
KEYWORDS
from an experienced construction consultant is required to estimate construction costs Deep learning; construction
accurately. Machine learning (ML) technology can learn from previous data, which is equivalent cost estimation; framework;
to human experience. Many project-specific ML models estimate the construction cost, which validator
misses the generalizability. This paper addresses the gap and designs, develops, implements,
and analyzes a deep learning (DL) based novel framework that maps 94.67% of the
independent variables with a mean average percentage error (MAPE) of 11.60%. The proposed
framework is not limited to any specific project. It estimates the construction cost of similar
projects, further validated by an innovative estimator validation unit.

Introduction
quantitative assessment, and thus, the prediction by
Constraining factors like costs, schedules, and quality the proposed system is more accurate.
all work together to ensure a project is completed suc- There has been a growing body of literature employing
cessfully. The construction phase provides an opportu- both classical statistical techniques (Akintoye & Fitzger-
nity to inspect and enhance a project’s quality while ald, 2000; Chan & Park, 2005; Hitsanu, 2022) and
ensuring that the schedule and cost remain within machine learning (ML) models (Alshboul et al., 2022b;
the bounds of its contracts. The contractors and the Kim & Cha, 2022; Matel et al., 2022; Shoar et al., 2022).
stakeholders are very involved in these estimates (Mar- The state-of-the-art machine learning approach is deep
kiz & Jrade, 2022). When stakeholders and decision- neural network (DNN)-based algorithms. This paper
makers have a realistic idea of how much a project uses DNN to estimate construction costs. The DNNs
will cost before it even begins, they can make informed are capable of establishing relations among complex, het-
decisions about feasibility studies, bidding, and cash erogenous, and multidimensional features, and so the
flow management (Al-Nassafi, 2022). To the detriment proposed methodology generates better and more
of the project’s stakeholders and contractors, cost over- reliable estimations. Sometimes the data can be more cru-
runs are a typical result of an underestimated project cial than the process itself (Amoore, 2022). However,
budget (Banks-Grasedyck et al., 2022). Several developing and optimizing machine learning models
methods have been used in practice, and others have for construction cost estimation in data-specific scenarios
been proposed in the literature to accurately predict demonstrates promising results, which have been studied
construction costs to limit losses and satisfy project and presented in this paper.
profitability targets. Multiple construction variables impact the construc-
Costs can be estimated in a few different ways, but tion cost directly and indirectly. However, they demon-
the two most common are qualitative and quantitative strate irregular patterns among themselves when the
assessments. Qualitative methods relying on experts’ cost is considered the target variable and related to
opinions may be biased, resulting in erroneous esti- these influencing variables (Elhegazy et al., 2022). More-
mates (Strömbäck & Tärnell, 2022). The proposed over, these variables are further dependent on other
methodology estimates construction costs from indirect variables (Okonkwo et al., 2022). As a result,

CONTACT Salman Saeidlou [email protected]

© 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http://creativecommons.org/licenses/by-nc-
nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built
upon in any way. The terms on which this article has been published allow the posting of the Accepted Manuscript in a repository by the author(s) or with their consent.
2 S. SAEIDLOU AND N. GHADIMINIA

the traditional simple summation of construction cost . Incorporate an estimation validator unit to address
variables is not enough to accurately estimate the cost. the credibility issue of practical uses of DNN in con-
The proposed framework overcomes this limitation. struction cost estimation.
The complex distribution pattern, multiple internal, . Analyze the limitation of the DNN-based construc-
external, and hidden dependencies, and their temporal tion cost estimator and the scope of mining opportu-
instability make construction cost estimation challen- nities from the limitations.
ging (Dang-Trinh et al., 2022). This challenge has
been beaten using the DNN-based construction cost
estimation framework presented in this paper. Usually,
Literature review
rule-based approaches are not enough for complex
relations between features and the target variable. The The calculation of building costs has always been a task
promising human-intelligence-like performance of that places a premium on the knowledge and insight of
deep learning (DL) technology is being used in this industry professionals (Elhag et al., 2005). That means
complicated field of studies, including: the medical ima- human intelligence with acquired experience can esti-
ging sector (Faruqui et al., 2021); stock market analysis mate construction costs. The DNN-based solution,
(Kumbure et al., 2022); targeted marketing (Sun et al., which exhibits human-like intelligence, is thus a good
2022); autonomous vehicles (Hui et al., 2022); virtual fit in the field explored in this paper. Not every organiz-
assistants (Liao et al., 2022); robotics (Nguyen et al., ation, especially small or newly established construction
2022); and in many other fields. Inspired by DL’s capa- firms, can afford to allocate a budget to keep experi-
bility in human intelligence replication, this paper enced consultants (Choudhry, 2016). Large organiz-
designs, studies, analyzes, and experiments, applying ations or companies may have the resources and
deep neural networks (DNN) in construction cost experience to compile their in-house construction cost
estimation. database. Apart from large construction firms, some
Despite the successful application of artificial neural individuals take construction responsibilities into their
networks (ANN) in construction cost estimation own hands, and those who aren’t construction industry
(Baduge et al., 2022), the coherence and relevancy of experts may rely on commercial vendors’ published
the recent literature are not leaning towards a standar- construction cost indexes (Zhang et al., 2017). Govern-
dized solution in this regard (Zabin et al., 2022). It is ment construction cost statistics also assist people in
a significant research gap in this sector, which has estimating construction costs. However, these data are
been observed and studied by the researcher of this not always up to date (K’akumu, 2007). An automatic
paper, and thus, the framework presented in this computerized system employed to estimate the con-
paper leads the way to standardize the methodology of struction cost, which can replace the necessity of using
applying machine-learning solutions to estimate con- a consultant or government statistics, benefits many
struction costs. This study proposes an innovative con- people. The research conducted in this paper is such
struction cost estimation framework using DNN to an endeavour.
bridge the gap. Another research gap is the applied vali- Like many different sectors, researchers in the con-
dation of the DNN-based cost estimator. It is a common struction sector have started using statistical and
practice in the deep learning domain to split the data set machine learning methods to improve the precision
into training, test, and validation sets and evaluate the and timeliness of cost estimates (Makridakis et al.,
validity of the network (Kahloot & Ekler, 2021). How- 2018). Statistical and machine learning methods
ever, construction cost estimation is an influential improve the decision-making process by transforming
business factor that directly impacts the stakeholders’ data from the past into decision-support systems (Lee
benefits (Doloi, 2013); thus, algorithmic and external et al., 2016). This can overcome the lack of data for pre-
validation is essential in order to rely on cost estimation cise estimation at the beginning of a project. Based on
by DNN practically. However, external validation is a this observation, the proposed framework focused on
significant research gap in the application of DNN in developing innovative network architecture to estimate
construction cost estimation. This research has evolved, construction costs precisely and validate the estimation
encompassing these two research gaps, and contributes with an external validation unit.
by carrying out the following: A study by Al-Momani (1996) builds an LR model
for building cost prediction with three project features
. Design, implement, analyze, and evaluate a DNN- as explanatory variables. This research gives an idea of
based framework to estimate construction costs the variables to consider for construction cost esti-
using direct variables. mation in the proposed methodology. The research
 BUILDING RESEARCH & INFORMATION 3

focused on the effect of public procurement law on One of the reasons behind choosing the machine
construction costs in Turkey and applied decision learning (ML) method for the framework tested and
tree (DT), support vector machines (SVM), and artifi- developed in this paper is that it captures complicated
cial neural networks (ANN). Information related to correlations between input and output without requir-
projects, such as start and end dates, geographic ing the specification of mathematical representations.
scope, and discount percentages, were used as inputs A systematic review by Tayefeh Hashemi et al. (2020)
(Erdis, 2013). Although this experiment aims to find discusses the common approaches to construction cost
the deviation from the estimated time and cost, it estimation using machine learning techniques, includ-
lays down an essential theoretical background to ing the support vector machine (SVM), the dynamic
adopt in developing the construction cost estimator tree (DT), and the random forest (RF). The SVM then
framework. separates the data along a hyperplane by nonlinearly
To better anticipate the costs of building in China, mapping the raw data into a high-dimensional space.
Shutian et al. (2017) created a fusion method that com- Using a recursive partitioning procedure, the DT can
bines the Kalman filter with least-squares support vector find a good tree structure inside a data set without
machines (LS-SVM) and linear regression (LR). The requiring expert knowledge (Tayefeh Hashemi et al.,
output of the experiment is promising. However, the 2020). However, the deep neural network approaches
variable distribution of the construction cost is non-lin- perform better than any other machine learning
ear. This raises the question of using linear approaches. approach in construction cost estimation (Wang et al.,
In the proposed methodology, the DNN has been used 2022). Comparing the effectiveness, robustness, opti-
to address this issue. Sub-gradient SVM has been used mizable nature, and capability to map between target
to evaluate the network’s performance along with LR. and complexly distributed dependent variables, the
There are many variables to estimate the construction DNN has been used as the cost estimator in the pro-
cost. Using those which are most influential in training, posed framework (Yanik et al., 2022).
a DNN is essential. The construction area, application
type, city hierarchy, and other project characteristics
Methodology
were used as inputs. The unit cost of concrete, the
unit cost of formwork, the type of structural assembly, The framework consists of a data set manager, a deep
and the amount of superimposed load were all taken neural network (DNN), and an estimation validator.
into account by Chakraborty et al. (2020) when estimat- These significant components have sub-components.
ing the total construction cost. A filtered version of these The overview of the proposed framework is illustrated
variables has been used in the proposed framework. in Figure 1.
An attempt to increase the precision of BIM labour
cost predictions was made by Huang and Hsieh (2020)
Data set preparation
by coupling random forest and linear regression. The
apartment complex’s total square footage and the total The proposed framework uses a specific data-set vari-
number of stories are part of the input variables. How- able pattern. It is essential to follow this pattern to get
ever, this approach is specific to a particular scenario. an optimized and accurate output from it. Table 1 lists
The proposed framework is not limited to a particular and explains the variables. These variables are common
problem but instead has a more straightforward and in every building construction. Depending on the
lightweight network architecture. It aims to develop a requirements, construction types, and specifications,
uniform platform for estimating construction costs. additional variables come into consideration. However,
An innovative construction cost estimation method the variables used in this paper are the fundamental
based on statistical analysis, particularly regression features.
analysis, was proposed by Li et al. (2022), and it studied The core variables related to constructions are con-
the norm. A similar approach was published by Lowe sidered in this paper. Other variables influence the
et al. (2006) but with much more simplicity, clarity, cost (Bernagros et al., 2021). However, these have
and familiarity with the applied algorithms. Although been ignored for the simplicity of the framework.
these two methods are good construction cost estima-
tors, they require a linear data set that maintains math-
Data processing
ematical linearity. Compared to these approaches, the
performance of the proposed methodology is invariant The ranges of the variables are not uniform, but it is
to data set linearity because of the use of a DNN. Other- essential to transform them into a uniform scale. In
wise, the performance tends to degrade. this experiment, the mean normalization has been
4 S. SAEIDLOU AND N. GHADIMINIA

Figure 1. Overview of the proposed framework.

Table 1. Variable list, description, and pattern. testing, and validation. The literature review on the
Variable Role Description state-of-art machine learning approaches suggests that
O Cost This is the final estimated cost. the training and test ratio of 70:30 is a standard data
x1 Building type Different types of buildings, categorical splitting ratio (Presnell & Alper, 2019). Thus, the same
type. Here 0 = residential, 1 = 0
commercial ratio has been used in this paper.
x2 Number of rooms Numerical number of floors, including the
per floor ground floor
x3 Number of special Three facilities have been considered in
facilities this framework Categorial type. Here, 0 Subset of the data set
= no, 1 = lift, 2 = garage, 3 = both 1 and
2 A subset of the data set after processing using Equation
x4 Total floor area Total area, including all floors; the (1) is listed in Table 2. Except for building type and the
numerical value measured in square
metres number of special facilities, the rest of the features are
x5 Number of levels Number of floors, including the ground scaled between 0 and 1. The x1 and x3 are encoded
floor
x6 Floor area per level Area of individual floors using the one hot encoding method. The rest of the
x7 Construction Average construction worker cost of a values are directly used during the training and testing
worker cost project
period.
The subset of the data set presented in Table 2 has
been selected randomly, which gives a general idea
used to scale the variable values between 0 and 1.
about the overall characteristics of the complete data set.
Equation (1) has been used for mean normalization
(Saranya & Manikandan, 2013).
xi − m Deep neural network (DNN) architecture
oi = (1)
max(xi ) − min(x) This experiment uses a deep neural network with four
Here, oi is the normalized value of the ith feature. The m hidden layers to estimate the construction cost. A fully
is the simple mean. The mean of the categorical values connected network architecture with 28 hidden nodes
has not been used. Equation (1) applies to numerical has been used (Bird et al., 2019). It is illustrated in
variables only. Figure 2.
The network has seven input nodes, one output node,
and 28 nodes in each hidden layer. The output vector of
Data cleaning and splitting
Irrelevant observations, structural errors, outliers, and Table 2. Subset of the data set.
missing values severely impact the overall performance x1 x2 x3 x4 x5 x6 x7 O
of machine-learning models. The proposed framework 1 0.50 0 0.90 0.75 0.81 0.65 0.72
1 0.78 0 0.91 0.79 0.85 0.77 0.85
requires a clean data set. The experimenting data set 0 0.85 0 0.85 0.64 0.49 0.92 0.89
has been manually cleaned in this research. The pro- 1 0.90 3 0.81 0.81 0.55 0.75 0.91
0 0.80 1 0.79 0.67 0.61 0.75 0.90
posed framework splits the data set into training,
 BUILDING RESEARCH & INFORMATION 5

Figure 2. Deep neural network architecture.

the network is defined using Equation (2). output nodes, respectively, in this range. The adaptive
moment estimation (ADAM) optimizer (Kingma &
K l = sl (Bl + W l K l−1 ) (2) Ba, 2014) has been used in this paper to optimize
Here the K l is the output vector; Bl is the bias vector; W l the learning process.
is the weight matrix; and sl is the activation function.
The tanh, ReLU (Agarap, 2018), and sigmoid functions Estimation validation unit
have been used as the activation functions of the input
layer, hidden layers, and output layers, respectively, An innovative estimation validation unit introduces
and are defined by Equations (3), (4), and (5) sequen- uniqueness to the proposed framework. This unit has
tially. been created using four different machine-learning
algorithms to validate the estimation done by the
ex − e−x
tanh(x) = (3) DNN. These two machine learning models are linear
ex + e−x regression (LR) (Weisberg, 2005) and support vector
k(x) = max (0, x) (4) machines (SVM) (Hearst et al., 1998) with a sub-gradi-
ent descent algorithm (Shalev-Shwartz et al., 2011).
k(x) = x (5) The LR model consists of eight variables (xn ) with
The bias for active neurones is set to 1, and dropped- eight different learning parameters. The LR mode is
off neurones are 0 (Mianjy et al., 2018). defined by Equation (7). It discovers the linear relation
between the price estimation and construction cost esti-
mating variables.
y = u0 x0 + u1 x1 + . . . + u7 x7 + 1 (7)
Training and optimization
Here, the y is the estimated price; u0 = 1; x0 = 1; u1 to
The proposed framework has experimented with the
u7 are the learning parameters of x1 to x7 respectively.
Residential Building Data Set publicly available at
The 1 is the error term.
the Machine Learning Repository of the University
The variable distribution for building cost estimation
of California, Irvine (UCI) (Asuncion & Newman,
demonstrates a non-linear pattern in different segments.
2007). The initial weight plays a role in learning
To address this issue, the SVM with a sub-gradient des-
optimization, which has been done using the normal-
cent algorithm has been used in this experiment to vali-
ized Xavier weight initialization (Datta, 2020) defined
date the estimation by the DNN. The soft-margin-based
by Equation (6).
SVM (Hu et al., 2010) used in this paper tries to mini-
 














6 6 mize the expression of Equation (8).
W i = PD − , (6)  
n+m n+m 1 n
f (W, b) = max (0, 1 − yi (W xi − b))
T

Here, Wi is the initial weight; PD a uniform prob- n i=1









6 + lW 2 (8)
ability distribution between the range − and
n+m







Here, the f (W, b) is the convex function of weight
6
. The n and m are the numbers of input and matrix W and bias b. The construction variable
n+m
6 S. SAEIDLOU AND N. GHADIMINIA

distribution does not follow a specific scale. Equation Table 3. Performance on UCI data set.
(8) does not scale with n in iterations, which makes it Model R2 RMSE MAE MAPE
a good fit for construction cost estimation. DNN 0.96 27.94 17.22 11.60%
LR 0.82 34.10 25.09 16.99%
SVM 0.91 31.44 21.38 13.92%

Result and experimental evaluation

The proposed framework estimates the construction Table 4. Performance on reinforced concrete building data set.
Model R2 RMSE MAE MAPE
cost using a deep neural network and validates the result
DNN 0.969 23.42 14.50 9.81%
using SVM and LR. Each of these three algorithms per- LR 0.843 31.65 21.72 14.20%
forms regression in this context. The literature review SVM 0.930 27.11 19.76 11.52%
shows that the state-of-art evaluation matrices for
regression are the coefficient of determination (R2 )
(Ozer, 1985) defined by Equation (9); the root mean approaches. However, the attempts failed because of
square error (RMSE) (Chai & Draxler, 2014) expressed the unavailability of the experimenting data set and
in Equation (10); the mean of absolute error (MAE) different evaluation matrices of measurement. As a
(Chai & Draxler, 2014) formulated in Equation (11); result, the experimenting data sets have been merged
and the mean average percentage error (MAPE) (Good- and augmented by introducing random noise to the
win & Lawton, 1999) demonstrated in Equation (12). normalized values. The range of the noise is between
m 0.01 and 0.2. The augmented data set’s experimental
(ai − pi )2 results have been listed in Table 5.
R2 = 1 − m i=1 (9)
i=1 (ai − mean(a))
2
The experimental results demonstrate that the per-





















formance of the framework is satisfactory. The R2 for
1  m
the UCI, RCB, and augmented data sets are 0.96,
RMSE = × ( pi − ai )2 (10)
m i=1 0.969, and 0.91, respectively. That means the model
properly fits 94.67% data.
1 m
MAE = × | pi − ai | (11)
m i=1
Performance evaluation
1 
m
pi − ai The coefficient of determination (R2 ) value for the DNN
MAPE = × (12)
m i=1
ai for the three experimenting data sets is illustrated in
Figure 3. It shows a minor variation for the three data sets.
Here, the ai is the target variable provided in the data The R2 value for DNN and SVM are similar. The LR
set; pi is the corresponding predicted target variable, and works best for linear relations. The data set is not
m is the number of instances in the data set.

Experimental result Table 5. Performance on augmented data set.

Model R2 RMSE MAE MAPE
The proposed framework has experimented with two DNN 0.91 29.44 20.63 13.54%
different data sets and one augmented data set. These LR 0.78 36.83 28.87 18.01%
SVM 0.88 33.58 23.42 15.26%
data sets are the Residential Building Data Set developed
by the University of California, Irvine (UCI) (Asuncion
& Newman, 2007), the Reinforced Concrete Building
Data Set prepared by M. Y. Cheng and Hoang (2018),
and an augmented combination of these two data sets.
Table 3 lists the performance of the proposed frame-
work on the UCI data set.
The performance of the framework on the
Reinforced Concrete Building (RCB) Data Set is listed
in Table 4. The data set’s relevant variables that align
with the proposed framework have been used in this
experiment.
Multiple attempts have been made during the exper-
iment to compare the results with other similar Figure 3. Coefficient of determination (R2 ) evaluation.
 BUILDING RESEARCH & INFORMATION 7

linear. As a result, the R2 for the LR is lower than for four papers have been compared with the proposed fra-
the DNN and SVM. However, the average of R2 is mework and listed in Table 6.
94.67%, which indicates models properly fit the data The accuracy of the proposed construction cost esti-
distribution. The mean average percentage error mation framework using DNN and the validation unit is
(MAPE) has been illustrated for the three experiment- 94.67% which is higher than three of the comparison
ing data sets in Figure 4. papers. However, the result obtained by Hashemi et al.
The MAPE is a little higher for the LR, as expected. (2019) is 0.04% higher than the proposed framework,
However, the DNN and SVM exhibit similar MAPE which is a marginal difference. The methodology of Alex
values. The average MAPE for the DNN is 11.65%, et al. (2010) is much more complex than the proposed fra-
which demonstrates the correctness of the cost estimation. mework. The proposed method still holds the superiority
The average MAPE of the SVM is 13.56%. There is only a in terms of architectural simplicity, even if the accuracy is
1.91% difference between the MAPE of the DNN and the 0.04% lower than the work of Alex et al. (2010).
SVM. It indicates the estimation from the DNN is valid.
The experimental values show a similar nature to the
Limitations and future scope
MAE values, which have been illustrated in Figure 5.
The performance evaluation of the proposed frame- Any computerized system has limitations. The proposed
work highlights the validity of the estimation of the framework is not an exception. There are four main
DNN. The comparison of the three ML models with limitations of this framework, which have been dis-
three different data sets leaves no scope for questioning cussed here.
the correct estimation made by the framework.

Variable limit
Performance comparison The first limitation of the proposed framework is the
The performance of the proposed system has been com- variable limit. It has been designed with the most com-
pared to four similar methodologies. Each of these mon and influential seven building construction cost
approaches used different data sets and features. estimators. It cannot handle more than seven variables.
Because of the similarity in the methodology, these One of the key contributions of this research is obtain-
ing accurate results with a limited number of variables.
Moreover, a limited internal variation of the variables
has been used in the proposed methodology. A modern
residential building may have multiple unique features.
These features vary from building to building. As a
result, it is not possible to incorporate every unique fea-
ture a building may possess (Juszczyk, 2017). This is a
major limitation of the proposed framework and any
framework. At the same time, there are no hard and
fast rules limiting the number of construction variables,
which imposes another challenge that this paper has not
Figure 4. Mean average percentage error (MAPE) comparison. solved. It is another drawback related to the variable
limit. However, these variable limitations pave the way
for more research on this topic: to find the optimum
number of variables and systematically handle the
unique features of buildings. These opportunities will
be explored in subsequent papers.

Table 6. Performance comparison.

Paper Method Accuracy
Alshemosi and Alsaad Multifactor Linear Regression 92%
(2017)
Alex et al. (2010) Artificial Neural Network (ANN) 80%
Hashemi et al. (2019) ANN & Genetics Algorithm (GA) 94.71%
Hybridization
Rafiei and Adeli (2018) Machine Learning (ML) 89.90%
Proposed Deep Learning (DL) 94.67%
Figure 5. The MAE values of DNN, SVM, and LR.
8 S. SAEIDLOU AND N. GHADIMINIA

Data set structure heterogeneity the economic variables are not ignored either. It has
been observed that the framework becomes complicated
The heterogeneity of the data set structure imposes a
when both economic and non-economic variables are
challenge on an informed approach to estimating build-
considered. For the sake of simplicity, the economic
ing costs through a machine learning algorithm (Scheres,
variables are ignored in this study. However, research
2016). The methodology followed by the researchers
is ongoing by the author of this paper to develop
while developing the data sets varies based on the per-
another framework to estimate the impact of economic
spective, core focus, building type, time, location, and
variables on construction costs. In the future, these two
many other factors. Different data sets come up with
frameworks will be merged to prepare a complete con-
different structures depending on the point of interest,
struction cost estimation, including economic and non-
the emphasis of certain factors, or the goal of the data col-
economic variables.
lection project. As a result, applying a uniform frame-
A study conducted by Rafiei and Adeli (2018)
work becomes challenging. Designing, implementing,
included both economic and non-economic variables
and training a DNN based on a particular data set leaves
in construction cost estimation using machine learning.
no scope for complexity. However, comparing the
The proposed methodology loses its superiority to the
trained network with similar approaches becomes
research published by Rafiei and Adeli (2018) from
difficult because there is no available similar enough pub-
the variable diversity perspective. Green building con-
lished research on the same data set structure. This is a
struction is a recent and eco-friendly trend that is turn-
major limitation of the proposed framework for any suit-
ing into modern construction standards. The proposed
able methods. The data set deformation method, where
research does not take the green building construction
any construction cost-related data sets are dissolved
parameters into consideration; however, this was done
into the stream of fragmented data to reconstruct it
by Alshboul et al. (2022b). Considering the heavy con-
into a uniform structure by maintaining a standard pro-
struction equipment as one of the features, which was
portional range of values, is a potential solution to this
done by Alshboul et al. (2022a), would make the pro-
problem. However, the level of complexity of such an
posed methodology more reliable. However, these limit-
approach requires a separate study, which leaves scope
ations have not been overcome in the current state of
for another field of research in this domain.
the research.

Limited data repository Discussion and conclusion

The data are like fuel to machine-learning-based Accurate cost estimation is a vital step in project success.
approaches. Machine-learning models learn from pre- Usually, consultants with years of experience enjoy
vious data to estimate or predict the target variable. It handsome consultation fees to estimate construction
has been observed during this experiment that the data costs accurately. Today’s advanced machine-learning
repository for building construction cost estimation is algorithms, especially deep neural networks, learn
not enriched enough (Scheres, 2016). The literature from data and demonstrate human-like estimation
reviews suggest that most published literature uses UCI capability. However, applying deep learning and
data sets or privately collected data that are not publicly machine learning in construction cost estimation is
available. Moreover, the data collection methodologies still nominal. This research aimed to utilize the potential
are yet to be standardized. These challenges limit the of DNN to prepare a practical framework to estimate
capability of machine-learning algorithms to estimate construction costs. The implementation principle used
building construction costs. However, the future scope in this research mapped 94.67% of independent vari-
of the research, to develop a standard methodology for ables into target variables. Furthermore, it estimated
creating construction-cost-related data sets to build a the price with a MAPE of 11.60%. The estimator vali-
shared and rich data repository, which would accelerate dation unit validates the estimation from the DNN to
the machine-learning research in this sector is revealed. increase the reliability of the estimation. Whereas
most of the research focuses on project-specific cost
estimation, this paper took an innovative, generalized
Economic variables
approach to estimate the cost of any similar project.
The economic variables are indirectly valuable but A machine-learning model designed to predict the
directly impact the construction cost (Tas & Yaman, construction cost of a particular construction project
2005). The proposed framework’s limitation is that the is made redundant after the project is completed unless
economic variables have not been considered. However, a similar project is launched. Therefore, the
 BUILDING RESEARCH & INFORMATION 9

generalization addressed in this paper is significant in Architectural Management, 29(10), 4153–4176. https://
this research field. However, the nature of the problem doi.org/10.1108/ECAM-08-2020-0614
imposes multiple challenges in developing an all-in- Alshboul, O., Shehadeh, A., Almasabha, G., & Almuflih, A. S.
(2022b). Extreme gradient boosting-based machine
one framework; thus, the research scope of this paper learning approach for green building cost prediction.
was limited to residential buildings with seven cost esti- Sustainability, 14(11), 6651. https://doi.org/10.3390/
mators. However, this limitation opens the door to su14116651
further research. This paper is a milestone in applied Alshemosi, A. M. B., & Alsaad, H. S. H. (2017). Cost estimation
machine learning with an outstanding demonstration process for construction residential projects by using multi-
factor linear regression technique. Criterion, 6(6), 7. https://
of accurately estimating construction costs. It has been
doi.org/10.21275/ART20174128
observed in the literature review that the data processing Amoore, L. (2022). Machine learning political orders. Review
approaches are different in multiple papers. The data set of International Studies, 49(1), 1–17. https://doi.org/10.
used is also exclusive to unique construction projects. 1017/S0260210522000031
The methodology presented in this paper generalizes Asuncion, A., & Newman, D. (2007). UCI machine learning
the data processing approaches into a standard struc- repository.
Baduge, S. K., Thilakarathna, S., Perera, J. S., Arashpour, M.,
ture, potentially a state-of-the-art method to process Sharafi, P., Teodosio, B., & Mendis, P. (2022). Artificial
data for construction cost estimation. Making an impor- intelligence and smart vision for building and construction
tant business decision, such as large-scale construction, 4.0: Machine and deep learning methods and applications.
requires cross-validation. The existing construction cost Automation in Construction, 141, 104440. https://doi.org/
estimation methodologies estimate the costs. However, 10.1016/j.autcon.2022.104440
Banks-Grasedyck, D., Lippke, E., Oelfin, H., Schwaiger, R., &
those approaches’ external cross-validation of the pre-
Seemann, V. (2022). The underestimated success factor:
dicted cost is absent. The proposed construction cost People, in: Successfully managing S/4HANA projects (pp.
estimation framework not only predicts the target vari- 125–176). Springer.
ables with 94.67% accuracy but also validates the predic- Bernagros, J. T., Pankani, D., Struck, S. D., & Deerhake, M. E.
tion to enhance the acceptability of the projection. (2021). Estimating regionalized planning costs of green
infrastructure and low-impact development stormwater
management practices: Updates to the US environmental
Disclosure statement protection agency’s national stormwater calculator.
Journal of Sustainable Water in the Built Environment, 7
No potential conflict of interest was reported by the authors. (2), 2. https://doi.org/10.1061/JSWBAY.0000934
Bird, J. J., Ekárt, A., Buckingham, C. D., & Faria, D. R. (2019,
July). Evolutionary optimisation of fully connected artificial
References neural network topology. In Intelligent Computing-
Proceedings of the Computing Conference (pp. 751–762).
Agarap, A. F. (2018). Deep learning using rectified linear units Springer.
(relu). arXiv preprint arXiv:1803.08375. Chai, T., & Draxler, R. R. (2014). Root mean square error
Akintoye, A., & Fitzgerald, E. (2000). A survey of current cost (RMSE) or mean absolute error (MAE)?–arguments
estimating practices in the UK. Construction Management against avoiding RMSE in the literature. Geoscientific
and Economics, 18(2), 161–172. https://doi.org/10.1080/ Model Development, 7(3), 1247–1250. https://doi.org/10.
014461900370799 5194/gmd-7-1247-2014
Al-Momani, A. H. (1996). Construction cost prediction for Chakraborty, D., Elhegazy, H., Elzarka, H., & Gutierrez, L.
public school buildings in Jordan. Construction (2020). A novel construction cost prediction model using
Management and Economics, 14(4), 311–317. https://doi. hybrid natural and light gradient boosting. Advanced
org/10.1080/014461996373386 Engineering Informatics, 46, 101201. https://doi.org/10.
Al-Nassafi, N. M. (2022). The effect of cash flow variation on 1016/j.aei.2020.101201
project performance: An empirical study from Kuwait. The Chan, S. L., & Park, M. (2005). Project cost estimation using
Journal of Asian Finance Economics and Business, 9(3), 53– principal component regression. Construction
63. https://doi.org/10.13106/jafeb.2022.vol9.no3.0053 Management and Economics, 23(3), 295–304. https://doi.
Alex, D. P., Al Hussein, M., Bouferguene, A., & Fernando, S. org/10.1080/01446190500039812
(2010). Artificial neural network model for cost estimation: Cheng, M. Y., & Hoang, N. D. (2018). Estimating construction
City of Edmonton’s water and sewer installation services. duration of diaphragm wall using firefly-tuned least squares
Journal of Construction Engineering and Management, support vector machine. Neural Computing and
136(7), 745–756. https://doi.org/10.1061/(ASCE)CO.1943- Applications, 30(8), 2489–2497. https://doi.org/10.1007/
7862.0000184 s00521-017-2840-z
Alshboul, O., Shehadeh, A., Al-Kasasbeh, M., Al Mamlook, R. Choudhry, R. M. (2016). Appointing the design consultant as
E., Halalsheh, N., & Alkasasbeh, M. (2022a). Deep and supervision consultant on construction projects. Journal of
machine learning approaches for forecasting the residual Legal Affaires and Dispute Resolution in Engineering
value of heavy construction equipment: A management Construction, 8(4), 04516005. https://doi.org/10.1061/
decision support model. Engineering, Construction and (ASCE)LA.1943-4170.0000195
10 S. SAEIDLOU AND N. GHADIMINIA

Dang-Trinh, N., Duc-Thang, P., Nguyen-Ngoc Cuong, T., & autonomous vehicle trajectory prediction and correction
Duc-Hoc, T. (2022). Machine learning models for estimat- model. Physica A: Statistical Mechanics and its
ing preliminary factory construction cost: Case study in Applications, 593, 126869. https://doi.org/10.1016/j.physa.
southern Vietnam. International Journal of Construction 2022.126869
Management, 1–9. https://doi.org/10.1080/15623599.2022. Juszczyk, M. (2017). The challenges of nonparametric cost
2106043 estimation of construction works with the use of artificial
Datta, L. (2020). A survey on activation functions and their intelligence tools. Procedia Engineering, 196, 415–422.
relation with xavier and he normal initialization. arXiv pre- https://doi.org/10.1016/j.proeng.2017.07.218
print arXiv:2004.06632. K’akumu, O. A. (2007). Construction statistics review for
Doloi, H. (2013). Cost overruns and failure in project manage- Kenya. Construction Management and Economics, 25(3),
ment: Understanding the roles of key stakeholders in con- 315–326. https://doi.org/10.1080/01446190601139883
struction projects. Journal of Construction Engineering and Kahloot, K. M., & Ekler, P. (2021). Algorithmic splitting: A
Management, 139(3), 267–279. https://doi.org/10.1061/ method for dataset preparation. IEEE Access, 9, 125229–
(ASCE)CO.1943-7862.0000621 125237. https://doi.org/10.1109/ACCESS.2021.3110745
Elhag, T. M. S., Boussabaine, A. H., & Ballal, T. M. A. (2005). Kim, J., & Cha, H. S. (2022). Expediting the cost estimation
Critical determinants of construction tendering costs: process for aged-housing renovation projects using a prob-
Quantity surveyors’ standpoint. International Journal of abilistic deep learning approach. Sustainability, 14(1), 564.
Project Management, 23(7), 538–545. https://doi.org/10. https://doi.org/10.3390/su14010564
1016/j.ijproman.2005.04.002 Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic
Elhegazy, H., Chakraborty, D., Elzarka, H., Ebid, A. M., optimization. arXiv preprint arXiv:1412.6980.
Mahdi, I. M., Aboul Haggag, S. Y., & Abdel Rashid, I. Kumbure, M. M., Lohrmann, C., Luukka, P., & Porras, J.
(2022). Artificial intelligence for developing accurate pre- (2022). Machine learning techniques and data for stock
liminary cost estimates for composite flooring systems of market forecasting: A literature review. Expert Systems
multi-storey buildings. Journal of Asian Architecture and with Applications, 197, 116659. https://doi.org/10.1016/j.
Building Engineering, 21(1), 120–132. https://doi.org/10. eswa.2022.116659
1080/13467581.2020.1838288 Lee, H., Chung, S. H., & Choi, E. J. (2016). A case study on
Erdis, E. (2013). The effect of current public procurement law machine learning applications and performance improve-
on duration and cost of construction projects in Turkey. ment in learning algorithm. Journal of Digital Convergence,
Journal of Civil Engineering and Management, 19(1), 121– 14(2), 245–258. https://doi.org/10.14400/JDC.2016.14.2.245
135. https://doi.org/10.3846/13923730.2012.746238 Li, Q., Guo, L., & Zhou, H. (2022). Construction quality evalu-
Faruqui, N., Yousuf, M. A., Whaiduzzaman, M., Azad, ation of large-scale concrete canal lining based on statistical
A. K. M., Barros, A., & Moni, M. A. (2021). Lungnet: A analysis. FAHM, and Cloud Model. Sustainability, 14(13),
hybrid deep-CNN model for lung cancer diagnosis using 7663. https://doi.org/10.3390/su14137663
CT and wearable sensor-based medical IoT data. Liao, S. W., Hsu, C. H., Lin, J. W., Wu, Y. T., & Leu, F. Y.
Computers in Biology and Medicine, 139, 104961. https:// (2022). A deep learning-based Chinese semantic parser
doi.org/10.1016/j.compbiomed.2021.104961 for the almond virtual assistant. Sensors, 22(5), 1891.
Goodwin, P., & Lawton, R. (1999). On the asymmetry of the https://doi.org/10.3390/s22051891
symmetric MAPE. International Journal of Forecasting, 15 Lowe, D. J., Emsley, M. W., & Harding, A. (2006). Predicting
(4), 405–408. https://doi.org/10.1016/S0169-2070(99)00007-2 construction cost using multiple regression techniques.
Hashemi, S. T., Ebadati E, O. M., & Kaur, H. (2019). A hybrid Journal of Construction Engineering and Management,
conceptual cost estimating model using ANN and GA for 132(7), 750–758. https://doi.org/10.1061/(ASCE)0733-
power plant projects. Neural Computing and Applications, 9364(2006)132:7(750)
31(7), 2143–2154. https://doi.org/10.1007/s00521-017- Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018).
3175-5 Statistical and machine learning forecasting methods:
Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., & Scholkopf, Concerns and ways forward. PloS One, 13(3), e0194889.
B. (1998). Support vector machines. IEEE Intelligent https://doi.org/10.1371/journal.pone.0194889
Systems and Their Applications, 13(4), 18–28. https://doi. Markiz, N., & Jrade, A. (2022). Integrating an expert system
org/10.1109/5254.708428 with BrIMS, cost estimation, and linear scheduling at con-
Hitsanu, M. S. (2022). Conceptual Cost Estimation of Highway ceptual design stage of bridge projects. International
Earthwork Construction in Iowa Using Spatial Statistical Journal of Construction Management, 22(5), 913–928.
Modeling. [Doctoral dissertation, North Dakota State https://doi.org/10.1080/15623599.2019.1661572
University]. Matel, E., Vahdatikhaki, F., Hosseinyalamdary, S., Evers, T., &
Hu, Q., Che, X., Zhang, L., & Yu, D. (2010). Feature evaluation Voordijk, H. (2022). An artificial neural network approach
and selection based on neighborhood soft margin. for cost estimation of engineering services. International
Neurocomputing, 73(10-12), 2114–2124. https://doi.org/ Journal of Construction Management, 22(7), 1274–1287.
10.1016/j.neucom.2010.02.007 https://doi.org/10.1080/15623599.2019.1692400
Huang, C. H., & Hsieh, S. H. (2020). Predicting BIM labor cost Mianjy, P., Arora, R., & Vidal, R. (2018, July). On the implicit
with random forest and simple linear regression. bias of dropout. In Jennifer Dy & Andreas Krause (Eds.),
Automation in Construction, 118, 103280. https://doi.org/ International conference on machine learning (pp. 3540–
10.1016/j.autcon.2020.103280 3548). PMLR.
Hui, F., Wei, C., ShangGuan, W., Ando, R., & Fang, S. (2022). Nguyen, H. T., Cheah, C. C., & Toh, K. A. (2022). An analytic
Deep encoder–decoder-NN: A deep learning-based layer-wise deep learning framework with applications to
 BUILDING RESEARCH & INFORMATION 11

robotics. Automatica, 135, 110007. https://doi.org/10.1016/ Strömbäck, A., & Tärnell, E. (2022). Evaluation and
j.automatica.2021.110007 Learning about Social Sustainability in the Real Estate
Okonkwo, C., Evans, U. F., & Ekung, S. (2022). Unearthing Industry: A Qualitative and Quantitative Study of how
direct and indirect material waste-related factors underpin- Real Estate Companies can Contribute to Society and
ning cost overruns in construction projects. International Profitability.
Journal of Construction Management, 1–7. https://doi.org/ Sun, C., Adamopoulos, P., Ghose, A., & Luo, X. (2022).
10.1080/15623599.2022.2052431 Predicting stages in omnichannel path to purchase: A
Ozer, D. J. (1985). Correlation and the coefficient of determi- deep learning model. Information Systems Research, 33
nation. Psychological Bulletin, 97(2), 307. https://doi.org/ (2), 429–445. https://doi.org/10.1287/isre.2021.1071
10.1037/0033-2909.97.2.307 Tas, E., & Yaman, H. (2005). A building cost estimation model
Presnell, K. V., & Alper, H. S. (2019). Systems metabolic based on cost significant work packages. Engineering
engineering meets machine learning: A new era for data- Construction and Architechtural Management, 12(3), 251–
driven metabolic engineering. Biotechnology Journal, 14 263. https://doi.org/10.1108/09699980510600116
(9), 1800416. https://doi.org/10.1002/biot.201800416 Tayefeh Hashemi, S., Ebadati, O. M., & Kaur, H. (2020). Cost esti-
Rafiei, M. H., & Adeli, H. (2018). Novel machine-learning mation and prediction in construction projects: A systematic
model for estimating construction costs considering econ- review on machine learning techniques. SN Applied Sciences,
omic variables and indexes. Journal of Construction 2(10), 1–27. https://doi.org/10.1007/s42452-020-03497-1
Engineering and Management, 144(12), 04018106. https:// Wang, R., Asghari, V., Cheung, C. M., Hsu, S. C., & Lee, C. J.
doi.org/10.1061/(ASCE)CO.1943-7862.0001570 (2022). Assessing effects of economic factors on construction
Saranya, C., & Manikandan, G. (2013). A study on normaliza- cost estimation using deep neural networks. Automation in
tion techniques for privacy preserving data mining. Construction, 134, 104080. https://doi.org/10.1016/j.autcon.
International Journal of Engineering Technology (IJET), 5 2021.104080
(3), 2701–2704. Weisberg, S. (2005). Applied linear regression (Vol. 528). John
Scheres, S. H. (2016). Processing of structurally heterogeneous Wiley & Sons.
cryo-EM data in RELION. Methods in Enzymology, 579, Yanik, E., Intes, X., Kruger, U., Yan, P., Diller, D., Van Voorst,
125–157. https://doi.org/10.1016/bs.mie.2016.04.012 B., … De, S. (2022). Deep neural networks for the assess-
Shalev-Shwartz, S., Singer, Y., Srebro, N., & Cotter, A. (2011). ment of surgical skills: A systematic review. The Journal
Pegasos: Primal estimated sub-gradient solver for svm. of Defense Modeling and Simulation, 19(2), 159–171.
Mathematical Programming, 127(1), 3–30. https://doi.org/ https://doi.org/10.1177/15485129211034586
10.1007/s10107-010-0420-4 Zabin, A., González, V. A., Zou, Y., & Amor, R. (2022).
Shoar, S., Chileshe, N., & Edwards, J. D. (2022). Machine learn- Applications of machine learning to BIM: A systematic lit-
ing-aided engineering services’ cost overruns prediction in erature review. Advanced Engineering Informatics, 51,
high-rise residential building projects: Application of ran- 101474. https://doi.org/10.1016/j.aei.2021.101474
dom forest regression. Journal of Building Engineering, 50, Zhang, S., Bogus, S. M., Lippitt, C. D., & Migliaccio, G. C.
104102. https://doi.org/10.1016/j.jobe.2022.104102 (2017). Estimating location-adjustment factors for concep-
Shutian, F., Tianyi, Z., & Ying, Z. (2017). Prediction of con- tual cost estimating based on nighttime light satellite ima-
struction projects’ costs based on fusion method. gery. Journal of Construction Engineering and
Engineering Computations, 34(7), 2396–2408. https://doi. Management, 143(1), 04016087. https://doi.org/10.1061/
org/10.1108/EC-02-2017-0065 (ASCE)CO.1943-7862.0001216