See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/329260095

Soft Sensor: Traditional Machine Learning or Deep Learning

Conference Paper · October 2018

DOI: 10.5151/2237-0234-31915

CITATIONS READS

0 1,904

5 authors, including:

Eduardo Magalhaes Oliveira

ihm stefanini
15 PUBLICATIONS 16 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Big Data with Advanced Analytics Features View project

SISTEMAS ESPECIALISTAS NEBULOSOS View project

All content following this page was uploaded by Eduardo Magalhaes Oliveira on 31 January 2019.

The user has requested enhancement of the downloaded file.

 SOFT SENSOR: MACHINE LEARNING TRADICIONAL OU
DEEP LEARNING?
Pablo Drumond1
Daniele Kappes2
Cassio de Moraes3
Eduardo Oliveira4
Mariana Teixeira5

Resumo
Para atingir a qualidade desejada no concentrado da flotação, numa usina de
mineração de ferro, o operador tomava decisões com latência mínima de 2 horas, o
equivalente ao tempo para recebimento de resultados da análise laboratorial. Um
tempo morto tão elevado, tornava difícil colocar a planta em regime quando ocorria
uma variação na alimentação da flotação ou na qualidade desejada. Dessa forma, o
objetivo do presente trabalho foi construir um sensor virtual capaz de estimar o teor
SiO2 e permitir a atuação do operador em um período mais curto. A alta variabilidade
na alimentação, adicionada às dificuldades inerentes da predição envolvendo séries
temporais tornaram o projeto singular. Foram aplicadas técnicas de mineração de
dados e seleção de features em uma base de 03 meses abrangendo variáveis da
Flotação e também da Deslamagem, e obtiveram-se entradas para elaboração de 03
modelos de Machine Learning: Random Forest, Gradient Boosted Trees e Multi Layer
Perceptron. O Soft Sensor baseado em redes neurais artificiais se mostrou
estatisticamente mais eficiente na etapa de teste. Na versão online, o baixo erro médio
absoluto obtido comprovou a robustez do modelo, entregou agilidade para a operação
e certificou o poder dessa abordagem em processos industriais com alta latência de
resultados laboratoriais.
Palavras-chave: Flotação; Machine Learning; Soft Sensor; Multi-Layer Perceptron

SOFT SENSOR: TRADITIONAL MACHINE LEARNING OR DEEP LEARNING?

Abstract
In an iron ore mining plant, in order to achieve the desired quality in the flotation
concentrate, the operator took decisions based on the result of a laboratory analysis
that took 2 hours minimum to be available. Such a long dead time made it difficult to
put the plant in steady state when a variation in the feeding of the flotation occurred or
when the output specification changes. Thus, the aim of the present work is to report
a construction of a virtual sensor capable of estimating the SiO2 content and that
allowed the operator to perform a responsive action in a shorter period. The high
variability in feeding on top of the inherent difficulties in the prediction involving time-
series made the project unique. Data mining techniques and feature selection were
applied on a 03 months database covering Flotation as well as the Desliming variables,
and different datasets were used in order to obtain the following three Machine
Learning models: Random Forest, Gradient Boosted Trees and Multi-Layer
Perceptron. The Soft Sensor based on artificial neural networks was more efficient in
the test stage. In the online version, the low average absolute error obtained proved
the robustness of the model, delivered agility to the operation and confirmed the power
of this approach in industrial processes with high latency of laboratory results.
Keywords: Froth Flotation; Machine Learning; Soft Sensor; Multi-Layer Perceptron.

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
1 Bachelor of Engineering (BEng) focused in Control and Automation from Federal University of Ouro
Preto, Graduated in Industrial Engineering and Data Science & Big Data from PUC-MG, Data
Engineer, Data Science, IHM Stefanini, Belo Horizonte, Minas Gerais, Brazil.
2 Bachelor of Engineering (BEng) focused in Control and Automation from Federal University of Minas
Gerais, Data Scientist, Data Science, IHM Stefanini, Belo Horizonte, Minas Gerais, Brazil.
3 Bachelor of Engineering (BEng) focused in Mechatronics from Federal Center of Technological
Education of Minas Gerais, Data Scientist, Data Science, IHM Stefanini, Belo Horizonte, Minas
Gerais, Brazil.
4 Master of Science in Electrical Engineering from Federal University of Minas Gerais, Leader of Data
Science and AI, Data Science, IHM Stefanini, Belo Horizonte, Minas Gerais, Brazil.
5 Bachelor of Statistics from Federal University of Minas Gerais.

1 INTRODUCTION

Froth Flotation is a mineral separation process based on two particles surface

characteristics: hydrophilicity and hydrophobicity. Hydrophobic particles adhere to the
air bubbles surface, have their density changed, then ascend and overflow the cell,
creating a froth, which is constantly removed from the cells. Hydrophilic particles
remain in the water, sediment and create an underflow with a higher concentration of
Fe, which flows into the next cell in order to refine the same process. Since some Fe
is carried outward along with the froth a closed circuit is needed to increase the metallic
recovery. This process is one of the most effective ways to take SiO 2 out of the slurry
and therefore increase the concentration of iron ore fines in the product stream. Figure
1 shows a froth flotation cell.

Figure 1 - Froth flotation cell with sensor, actuators and control system.
Source: https://www.researchgate.net [1]

In a typical froth flotation plant, the operator takes some actions, such as increasing
the reagents flow or reducing feed rate, in order to achieve the desired quality. In the
production facility studied, those actions can only have their effectiveness confirmed
by the laboratory analysis results, within 02 hours. Such a long dead time made it
difficult to put the plant in a steady state when feeding quality varies or when the output
specification changes.

Therefore, a possible solution for this issue, that will be referenced as SiO2 Prediction
Problem, would be a predictive model capable of estimating the concentration of SiO 2
in the product stream, which turned into the central objective of the project. The main
reason to support the investment is related to product quality, i. e., meeting the product

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
specification with lower resource consumption, shorter time and thus delivering a better
control of the plant with lower cost as opposed to the costly online slurry analyzers.

In broad terms, applying prediction in mineral processing, especially Froth Flotation, is

not a new concept. There are several types of research involving different machine
learning techniques, and it is worth mentioning the following as the being the ones
most relevant to this work:
 Shabahzi et al. [2] investigated the effect of process variables on the flotation
rate and recovery using Random Forest;
 Nakhaei and Irannajad [3] used various neural networks architectures in order
to predict the metallurgical performance of the flotation column in a copper plant;
 Despotoviç [4] built a Support Vector Regression model to estimate deinking
flotation performance in the paper recycling process.
Nevertheless, the majority of the previous works applied just train, validation and test
steps in order to prove the concept and they did not report deploying them nor
achieving online phase.

Among the challenges that made this paper pertinent, it is possible to mention the non-
existence of machine vision system, thus features like bubble size and froth speed
were not available to take part as model inputs. Furthermore, on account of being
essentially a dynamic process, there were no strong correlations between variables
and output, and traditional statistical methods were not efficient enough to describe the
relationship among variables, which required great efforts in feature engineering
phase.

2 METHODOLOGY

According to Mariscal et al. [5], Cross-Industry Standard Process for Data Mining,
known as CRISP-DM, is the most widely used form of data-mining model, and it was
also assured by KDnuggets research [6] to be the top methodology for data science
projects. Figure 2 shows the cycle and phases of the process.

Figure 2 - CRISP-DM Process Diagram

Source: https://www.kdnuggets.com

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
Despite having no enhancements since its release, even as big data advanced, it
proved to be a powerful approach for analytics and consequently was selected to guide
the development of the SiO2 prediction model.

It is important to mention that even though CRISP-DM proposes a cycling approach,

which means the developing model is enhanced within each loop, the task behind its
phases will be described in this paper as a single thread.

2.1 Business Understanding

A preliminary assessment on the product stream showed that variability of SiO 2 over a
relatively short period of time was very common, which indicated recurrent operator
attempts to meet product specification and confirmed the hypothesis that the flotation
plant is indeed difficult to operate. Figure 3 shows that variability.

Figure 3 - Concentration of SiO2 in the product stream

Source: Author’s own figure

Contrary to the initial belief, the objective is not to minimize SiO2 in the output stream.
In fact, blending iron ore and ending up with a product within the admissible range is
the major goal of a mineral processing plant. Strictly speaking, sometimes a poor slurry
delivered by flotation is advantageous to the business.

Thus, after the initial assessment, the goals were established as being: to find the root
causes of SiO2 variance; to understand the influence of certain variables such as
surfactants and wetting agents flow, and then, to build a model taking into account the
most important features for SiO2 estimation.

2.2 Data Understanding

From the start of the analysis, the PIMS showed itself to be the appropriate long-term
source for retrieving necessary data for modeling. It stores and aggregates not just
time series of process variables but also sampling results which are published by the
LIMS.

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
The initial dataset was 03 months long. At the first cycle, a primary concern was to
assure that any changes in the flotation circuit had not happened in the previously
mentioned period. It could occur mainly because most of the time a cell designed to
be the second stage of Scavenger is turned into a third Cleaner stage although without
any notification in the automation system.

In order to understand the plant performance, it would be interesting to include in the

data collection a high variance of ROM1. On top of that, there was another assumption
related to the variable extraction: the value must be available online, which means that
during runtime the model must have access to the current value of the variable
considered as input.

Finally, after meeting the above requirements, data were extracted and loaded into a
relational database in order to make it easy operations like select, filter and join.

Furthermore, it is important to highlight the sampling operation, by which the

concentration of SiO2 is determined in the flotation and deliming processes. The lab
sample, which is drawn from the process every 2 hours, consists of small partitions of
slurry that are collected at 12-minute intervals. Therefore, one lab result reflects only a
combination of those partitions and not the process scenario at a specific moment. The
mentioned sampling method may lead to considerable loss of information since some
process changes cannot be properly tracked during a 2-hour time frame.

2.3 Data Preparation

The dataset shape was 700k rows and 120 columns, which corresponds to 10 seconds
sampling of the process and lab variables. The number of rows was good enough to
avoid underfitting, nevertheless, the number of columns could impact in generalization
performance, which required extra efforts for selection and reduction.

Data preparation or pre-processing handles and transform raw data so that knowledge
could be easily discovered [7]. The following enrichment methods were performed:
 Purging plant shutdown periods based on feeding rate value;
 Filtering outliers with z-score and threshold adjustment;
 Deleting features with extreme low variance, for instance, some cell levels.

The correlation coefficient between variables and target was extremely weak, which
confirmed the initial hypothesis inherent to the dynamic process. However, it was worth
reducing some redundant features, for instance, densities and motor currents and air
insufflation rate among the cells.

The first dataset (DS#1) was prepared after applying Feature Ranking with Recursive
Feature Elimination, known as RFE. For this purpose, it was used a Support Vector
Regression algorithm as an external estimator in order to assign weights to each
feature. Then, the process variables were recursively pruned until the desired number
was reached, 20.

1
Run of Mine

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
The second dataset (DS#2) was prepared after applying Least Absolute Shrinkage and
Selection Operator. Lasso is a regression analysis method that performs both variable
selection and regularization in order to enhance the prediction accuracy and
interpretability of the statistical model it produces [8]. This method selected 41
variables using 0,025 as the regularization term.

It is relevant to note that, despite similar principles, the methods above elected quite
different features. Then, it was decided to prepare a third dataset (DS#3) made up of
common variables between the output of RFE with 40 features and Lasso with 41
features. So it was selected 26 process variables.

Although PCA is a notorious tool for dimension reduction, it was not carried out in this
study due to a loss of understanding of flotation process. In other words, it would be
difficult to interpret whether the transformed vectors were clear enough according to
the process knowledge.

2.4 Modelling

Regardless of machine learning technique to be implemented, the approach would rely

on three phases: training, validation and test. Each step counting on a data subset
from each the three major datasets would be prepared by the feature engineering, split
as follows:
 Test subset: last 12 days of the working dataset;
 Train subset: 80% of the remaining rows, randomly selected;
 Validation subset: the remaining rows.

In general, in the first step the model parameters are adjusted using an optimization
algorithm which aims to minimize error measurement, using the difference between
estimated and expected output, also called supervised learning technique.

The second step is applied to tune hyperparameters in order to overcome bias or

overfitting behaviors. It is followed by a comparison between the results of each
attempt and a selection of which best fits the data. Alternatively, it is possible to merge
Train and Validation subsets in order to use cross-validation technique then, choosing
the optimal weights among models trained against K-Folds.

Lastly, the performance of the trained model is assessed with unseen data, which
simulates a real context with a smaller dataset that follows the same probability
distribution of the training phase.

Gradient Boosted Tree, Random Forest and Multi-Layer Perceptron were chosen to
try to solve SiO2 Prediction Problem. Root Causes Analysis was carried out to better
understand the influence of some variables on the chemical concentration of the output
stream.

Froth Flotation is a dynamical process, which means that training the model with the
variables at the same instant of time would certainly result in a poor performance. In
other words, SiO2 at a specified timestamp is not related to the reagents flow in the

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
same timestamp, which was the reason one of the biggest challenges was to find out
the right lag between variables and the target.

2.4.1 Gradient Boosted Tree

Gradient Boosting is a machine learning technique that can be used for regression and
classification problems. It produces a prediction model in the form of an ensemble of
simple estimators, typically decision trees. In this method, the learning procedure
consecutively fits a new model to provide a more accurate estimate of response
variable [9].

2.4.2 Random Forest

Random Forest is a machine learning technique that can also be used for regression
and classification problems. It constructs a multitude of decision trees at training time
and outputting the mean prediction (regression) of the individual trees. “It consists of a
chosen number of decision trees. Each of the decision tree models is learned on a
different set of rows (records) and a different set of columns (describing attributes),
whereby the latter can also be a bit-vector or byte-vector descriptor (e.g. molecular
fingerprint). The row sets for each decision tree are created by bootstrapping and have
the same size as the original input table. For each node of a decision tree, a new set
of attributes is determined by taking a random sample of size sqrt(m) where m is the
total number of attributes. The output model describes a random forest and is applied
in the corresponding predictor node using a simple majority vote” [10].

2.4.3 Root Cause Analysis

Another approach used in this study was to try to build a root cause analysis hoping to
get the causes of the SiO2 percentage being out of specification. Two different
approaches were tried : (1) using process and quality variables without any lag and (2)
using this same dataset but applying a one-hour lag to the process variables due to
the silica’s residence time. Then, after deciding for the second approach, it was
necessary to do some data manipulation before applying the Decision Tree algorithm
to get a root cause analysis: it was classified with the label “meet the requirement”
those data which the silica percentage was between 2.9% and 3.9% and as “does not
meet the requirement” those which the silica percentage was less than 2.9% and
greater than 3.9%.

In the next step, the dataset was balanced, so data were filtered, leaving the same
amount of data classified as “meet the requirement” and classified as “does not meet
the requirement”. Then, the dataset was split in training set (86%) and test set (14%)
and it was applied the Decision Tree Algorithm to both datasets.

2.4.4 Multi-Layer Perceptron

The Perceptron was introduced by McCulloch and Pitts as a computational version of

a biological neuron, which can be understood as a weighted sum of the inputs used in
an activation function to compute an output, as shown in Figure 4.

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
 Figure 4 - Perceptron
Source: http://wwwold.ece.utep.edu/research/webfuzzy/docs/kk-thesis

Based on that abstraction, Minsky and Papert proposed perceptrons organized in

layers interconnected and confirmed Neural Networks with more than one hidden layer
as a powerful tool capable of solving nonlinear separation problems.

As it turned out, the network with only one hidden layer achieved a low performance,
R-squared was about 0,59 in the training phase. By using deep learning techniques
along with a graphical analysis of cost function, the topology dramatically changed,
and it allowed to improve R-squared to 0,95 in the same phase. Which seemed to be
promising, in fact, exposed an overfitting behavior when the test phase resulted in R-
squared about 0,2. Then, in order to address this issue as well as to avoid getting stuck
in a local minimum, the following actions were performed during the training:
 Adjusting regularization term: adds penalties to the coefficients;
 Applying drop-out: ignores a random fraction of nodes;
 Adjusting decaying rate: drops the learning rate as the training proceeds;
 Setting batch size: defines the number of samples considered in each batch.

Finally, the topology used to solve SiO2 prediction problem counted on 08 hidden
layers and up to 50 nodes in one of them. It used Rectified Linear Unit (ReLU) as
activation function and Backpropagation with Adam as the training algorithm.

The table below shows the process variables of DS#1, used as inputs in the network:

Table 1 – DS#1 Features

Inputs
Conditioning Tank - pH CF-05 Cell (Rougher) - Amine Flow
CF-01 Flotation Cell (Rougher) - pH Conditioning Tank - Starch Flow
Conditioning Tank - Density CF-03 Cell (Rougher) Level
Grinding Feeder - Current CF-05 Cell (Rougher) Level
Recirculation Pump - Current CF-13 Cell (Cleaner 2) Level
Recirculation Pump - Speed CF-21 Cell (Scavenger 2) Level
CF-05 Cell (Rougher) - Stirrer Current SiO2 - Output Stream (02h Lag)
CF-07 Cell (Cleaner 1) - Stirrer Current Al2O3 - Output Stream (02h Lag)
CF-11 Cell (Cleaner 2) - Stirrer Current Fe - Output Stream (02h Lag)
CF-15 Cell (Scavenger 1) - Stirrer Current Mn - Output Stream (02h Lag)
CF-17 Cell (Scavenger 1) - Stirrer Current P - Output Stream (02h Lag)

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
 CF-19 Cell (Cleaner 3) - Stirrer Current Conditioning Tank - Amine Flow
CF-09 Cell (Cleaner 1) - Amine Flow

3 RESULTS

3.1 Evaluation

The next topics show the best result using the test subset of each dataset (DS#1, DS#2
and DS#3) for each model.

3.1.1 Gradient Boosted Tree

The Gradient Boosted Tree Algorithm applied to the DS#3 presented a mean absolute
error of 1,141 and coefficient of determination of 0,19. The dynamic response is shown
in Figure 5:

Figure 5 - Dynamic response of Gradient Boosted Tree algorithm

Source: Author’s own figure

3.1.2 Random Forest

The Random Forest Tree Algorithm applied to the DS#3 presented a mean absolute
error of 1,145 and coefficient of determination of 0,273. The dynamic response is
shown in Figure 6:

Figure 6 - Dynamic response of Random Forest Tree algorithm

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
 Source: Author’s own figure

3.1.3 Root Cause Analysis

The DS#2 had a better performance, and, using it, it was evaluated an accuracy of
94,346% in the test set, and the following rules to “silica meet the requirement”,
described in Table2:

Table 2 – Root Cause Rules

Rules
Output Stream - Fe Concentration > 66,805%
Feeding Flow Rate <= 2607,182 m3/h
Conditioning Tank - Amina Flow <= 8,391 m3/h

Therefore, 78,4% of the data with the characteristics listed in Table 2 corresponds to
“silica meet the requirement”.

However, this analysis should be interpreted carefully. Firstly, because of one of the
rules: the amine flow. According to this rule, if the amine rate is lower than 8,931m³/h,
there is a high chance of the silica meet the specification (low silica percentage). But
generally, the greater the amine flow the lower the silica percentage. So, this specific
rule captured by this model could not be a cause, but a consequence: once the plant
operator sees the high percentage of silica, he tries to increase the amine flow hoping
that the silica percentage falls. And secondly, this analysis should be more explored
using different residence time because the value used (one-hour lag) could not have
been precise enough.

3.1.4 Multi-Layer Perceptron

The MLP applied to the DS#1 presented a mean absolute error of 0,7974 and R
Squared of 0,4966. The expected output and the estimated value (or its moving
average, just to show a cleaner dashboard) are depicted in Figure 7.

Figure 7 - Dynamic response of MLP algorithm

Source: Author’s own figure

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
3.2 Deployment

Mean Absolute Error was chosen as the suitable metric to define the model to be
deployed mainly because, as it is absolute, it was easy to compare it to the target
variable and to decide whether the model met the performance requirements. The
values for each algorithm is presented in Table 3.
Table 3 - Mean Absolute Error for each model

Model Mean Absolute Error

Gradient Boosted Tree 1,141
Random Forest 1,145
Multi-Layer Perceptron 0,7974

Based on the results achieved it was decided to deploy the MLP model. The application
architecture built to estimate SiO2 online is depicted in Figure 8. It has two main python
scripts OPC Collector and Soft Sensor App. The first one is responsible for retrieving
current values of process variables every 20 seconds from plant floor via OPC and
store them in a relational database. The second one takes the last 05 minutes records,
also including lab results, estimate silica percentage using weights and bias computed
during the training phase and, finally, write the prediction to the PLC (the CPU that
controls the industrial process).

Figure 8 – Application Architecture

Source: Author’s own figure

4 CONCLUSION

After application setup, the plant operator was able to take actions every 05 minutes
relying on a prediction of SiO2 with 0,8% of mean absolute error. Compared to the
previous dead time, it was considered the great achievement of the project. And thus,
it certified the power of this approach in industrial processes with high latency of
laboratory results.

Some experiments showed that lab results related to the recirculation sampling
significantly enhanced the performance of all the models. However, it was discarded

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
 because the sampling process is carried out manually and thus, it is not considered
reliable enough to take part as model input. Then, the secondary achievement of the
study was to support the adoption of automated samplers on that stream.

As a future study, it was proposed to build and evaluate classification instead of

regression models, by discretizing the value of SiO2, which would widen the range of
possibilities and could lead to better results.

REFERENCES

1 Kramer A., Gaulocher S., Martins M., Leal Filho L. (2012) Surface Tension Measurement
for Optimization of Floating Control. Procedia Engineering. 46. 111–118.
10.1016/j.proeng.2012.09.453.
2 B. Shahbazi, S. Chehreh Chelgani, S.S. Matin (2017) Prediction of froth flotation
responses based on various conditioning parameters by Random Forest method.
Colloids and Surfaces A 529 (2017) 936–941
3 Nakhaei F., Irannajad M. (2015) Application and comparison of RNN, RBFNN and MNLR
approaches on prediction of flotation column performance. International Journal of Mining
Science and Technology 25 (2015) 983–990
4 Despotović V., Trumić M. S., Trumić M. Ž (2017) Modeling and prediction of flotation
performance using support vector regression. Recycling and Sustainable Development
10 (2017) 31-36
5 Mariscal G., Marbán Ó., Fernández C. (2010). A survey of data mining and knowledge
discovery process models and methodologies. The Knowledge Engineering
Review, 25(2), 137-166. doi:10.1017/S0269888910000032.
6 Piatetsky G. (2014) CRISP-DM, still the top methodology for analytics, data mining, or
data science projects. Available at: https://www.kdnuggets.com/2014/10/crisp-dm-top-
methodology-analytics-data-mining-data-science-projects.html. Accessed: 06/26/2018.
7 Pyle D. (1999). Data preparation for data mining. São Francisco: Morgan Kaufmann apud
De Casto L. N., Ferrari D. G. (2016). Introdução à Mineração de Dados. Editora Saraiva.
8 Lasso (Statistics). Available at https://en.wikipedia.org/wiki/Lasso. Accessed: 06/26/2018
9 Dhiman V., Bedi R., Kumar M. (2017) Predictive Big Data Analysis for Machine Learning
using Knime to solve many business challenges. Accessed: 06/26/2018. Available at:
http://www.innovativeresearchpublication.com/singapore/paper%2045.pdf
10 Random Forest Learner – NodePit (2017) KNIME. Accessed: 06/06/2018. Available at:
https://nodepit.com/node/org.knime.base.node.mine.treeensemble.node.randomforest.l
earner.classification.RandomForestClassificationLearnerNodeFactory

* Contribuição técnica ao 22° Seminário de Automação e TI, parte integrante da ABM Week, realizada
de 02 a 04 de outubro de 2018, São Paulo, SP, Brasil.
View publication stats