Hindawi

Computational Intelligence and Neuroscience

Volume 2021, Article ID 5572781, 19 pages
https://doi.org/10.1155/2021/5572781

Research Article
A New Random Forest Algorithm Based on Learning Automata

Mohammad Savargiv ,1 Behrooz Masoumi ,1 and Mohammad Reza Keyvanpour2

1
Faculty of Computer and Information Technology Engineering, Qazvin Branch, Islamic Azad University, Qazvin, Iran
2
Department of Computer Engineering, Alzahra University, Tehran, Iran

Correspondence should be addressed to Behrooz Masoumi; [email protected]

Received 12 February 2021; Revised 9 March 2021; Accepted 16 March 2021; Published 27 March 2021

Academic Editor: Nian Zhang

Copyright © 2021 Mohammad Savargiv et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
The goal of aggregating the base classifiers is to achieve an aggregated classifier that has a higher resolution than individual
classifiers. Random forest is one of the types of ensemble learning methods that have been considered more than other ensemble
learning methods due to its simple structure, ease of understanding, as well as higher efficiency than similar methods. The ability
and efficiency of classical methods are always influenced by the data. The capabilities of independence from the data domain, and
the ability to adapt to problem space conditions, are the most challenging issues about the different types of classifiers. In this
paper, a method based on learning automata is presented, through which the adaptive capabilities of the problem space, as well as
the independence of the data domain, are added to the random forest to increase its efficiency. Using the idea of reinforcement
learning in the random forest has made it possible to address issues with data that have a dynamic behaviour. Dynamic behaviour
refers to the variability in the behaviour of a data sample in different domains. Therefore, to evaluate the proposed method, and to
create an environment with dynamic behaviour, different domains of data have been considered. In the proposed method, the idea
is added to the random forest using learning automata. The reason for this choice is the simple structure of the learning automata
and the compatibility of the learning automata with the problem space. The evaluation results confirm the improvement of
random forest efficiency.

1. Introduction characteristics, namely, prioritization of features, attribution

of different weight coefficients to different classes, and il-
Random forest is one of the methods of ensemble learning lustration and unsupervised learning ability.
that comes under the homogeneous base learner category in According to the literature, the random forest method is
terms of the type of constructive classifiers. As the name one of the most practical methods of ensemble learning.
implies, all base learners are decision trees, and therefore Weighting the base learners in ensemble learning is one of
they have a simpler structure than similar methods [1]. The the main challenges in aggregating the basic classifiers in
random forest structure has two advantages. The first cat- order to achieve a stronger classifier [3]. The reason for
egory is from a computational point of view, and the second weighing base learners, or in other words, determining the
category is from a statistical point of view. Advantages that impact factor for each base learner, is to increase the scal-
can be considered from a computational point of view are: ability of the data mining algorithm with the problem space.
the random forest has the ability to deal with both regression This becomes even more apparent when the environment is
and classification issues. The train and prediction processes dynamic, and different or sometimes contradictory behav-
in this classifier are performed at high speed, and therefore iours are observed from data in different situations. The text
the random forest is known as one of the fast classic clas- data environment has such an interesting behaviour that it
sifiers. Another advantage of the random forest is its ability challenges data mining algorithms. For example, placing one
to be used directly in high-dimensional issues [2]. The word on one domain may create a positive polarity, but it
advantages of the second view of the random forest are its may also create a negative polarity on another domain. This
2 Computational Intelligence and Neuroscience

difference in polarity is created without any change in the In this paper, a brief review of random forest in terms of
form of the word and without any change in the role of the application scope is given.
word from a grammatical point of view. The word “small” in In this paper, a learning automata-based method is
both the electronic domain and the restaurant domain has proposed to improve the random forest performance.
such a behaviour. This behaviour poses a major challenge to
The proposed method operates independently of the
the opinion mining algorithms [4].
domain, and it is adaptable to the conditions of the
The classical solution in the literature to overcome this
problem space.
challenge is based on the use of lexical-based approaches.
This approach is based on frameworks such as unigram, The rest of the paper is organized as follows. In Section 2,
n-gram, aspect-based, and similar methods, and all of them related work is introduced. Section 3 presents the intro-
are data-dependent. In addition to the urgent need for duction to learning automata. The proposed method is
predefined data, these methods lose their efficiency if they explained in Section 4. Section 5 includes evaluation. Dis-
are met with an unspecified word or metaphor in the cussion is given in Section 6, and finally, the conclusion and
opinion mining field. In other words, they are not com- future work are described in Section 7.
patible with the problem space. The way random forest
works is that with the sequential placement of training data
and feature vectors that are injected into each of the base 2. Related Work
learners, it tries to find the best subset of features, and by
increasing their impact factor in the classifier, it achieves the In this section, theories and literature on the subject of
highest performance among all the aggregated base learners random forest are examined. The purpose of this section is to
[5]. However, this method is not effective in relation to data review the innovations that have been introduced around
such as text, in which a word can have different polarities in random forest in recent years.
different domains because, in the classification algorithm, Random forest is considered as one of the methods of
there is no ability to adapt to the conditions of the problem ensemble learning in the homogeneous ensemble learning
space. subgroup. In the random forest, each decision tree, or in
In this paper, we intend to empower random forest with other words, each base learner, has access to a random subset
the idea of reinforcement learning and improve its efficiency. of feature vectors [6]. Therefore, the feature vector is defined
In the proposed method, learning automata is used to ag- as follows:
gregate and weigh base learners. The way learning automata
works is to receive feedback from the environment and x � 􏼐x1 , x2 , ..., xp 􏼑, (1)
perform one of the actions based on the type of feedback. In , where p is the dimension property of the available vector for
the learning automata, feedbacks are divided into two cat- the base learner. The main goal is to find the prediction
egories of reinforcement signals: reward signals and penalty function as f(x) that predicts the Y parameter. The prediction
signals. For each reinforcement signal received by the function is defined as follows:
learning automata, it updates the probability of selecting the
selected action in the previous step. This process continues L(Y, f(x)), (2)
until the probability of action selections converges to one of
the actions; in other words, the best option for running in the where L is known as the loss function, and the goal is to
current situation is found. In the proposed method, learning minimize the expected value of the loss. For regression
automata actions are appropriate when one of the base applications and classification applications, squared error
learners selected leads to the maximum reward that can be loss and zero-one loss are common choices, respectively.
received from the environment. Since at each stage of These two functions are defined as follows in equations (3)
learning automata execution, the learning algorithm tries to and (4), respectively.
select the best option, achieving global optima in the
problem space is guaranteed. This is proof of the adaptability L(Y, f(x)) � 􏼐Y − f(x)2 􏼑, (3)
of the proposed method. In the proposed method, the
subprocess of replacing features in the feature vector is 0, if Y � f(x),
L(Y, f(x)) � I(Y ≠ f(x)) � 􏼨 (4)
removed, and all the features in the feature vector are used. 1, otherwise.
As a practical application in the field of opinion mining, if
the Bag of Word (BoW) method is used to create the feature To create an ensemble, a set of base learners come to-
vector, the advantage of considering all the features of the gether. If base learners are defined as follows:
feature vector will also cover cases that occur rarely. In other h1 (x), h2 (x), . . . , hJ (x), (5)
words, in the proposed method, the aspect of independence
from the domain in the processes such as opinion mining is for regression applications, the averaging will be based on
considered. equation (6), and for classification applications, the voting
Our contribution is summarized as follows: will be based on equation (7).
Computational Intelligence and Neuroscience 3

1 J use an oblique hyperplane to split the data for increasing the

f(x) � 􏽘 h (x), (6) accuracy of the trees and reduce the depth of RF. Probst et al.
J j�1 j
[23] tune the hyper-parameters to achieve higher perfor-
mance to improve the RF. Kim et al. [24] propose a method
J for interpreting and simplifying a black-box model of a deep
f(x) � arg max 􏽘 I􏼐y � hj (x)􏼑. (7) RF by quantifying the feature contributions and frequency of
j�1
the fully trained deep RF. Jain et al. [25] purpose dynamic
The Random Forest pseudocode for classification ap- weighing scheme for RF using the correlation between
plications is shown in Algorithm 1. decision tree and data samples. In the global problem field,
As can be seen in Algorithm 1, in the random forest, an Stafoggia et al. [26] estimate daily particulate matter for
attempt is made to find a subset of features using the various weather forecasting by RF. Modeling the global forest area by
replacements of training data and features that maximize the RF is the main target of [27]. Breidenbach and Saravi [28]
efficiency and accuracy of the output. This set of features is present research on land-subsidence spatial modeling and its
used to identify a new instance. assessment. Analyzing the net ecosystem carbon exchange is
The following is a brief review of the random forest the goal of [29]. Prediction about the global climate problem
subject literature. It should be noted that we intend to in- using the index quantization ability of random forest and the
troduce the background of the subject, and this paper is not a optimizing ability of PSO in the NN prediction model is the
review paper, and the presented review is a brief review and main purpose of [30]. Li et al. [31] solve the class imbalance
does not mention all the previous works undoubtedly. by detecting serial case pairs.
However, the authors have tried to refer to the latest and
most authoritative research work published in the recent 2.3. Healthcare field. Diagnosis detection and prediction of
years. obesity in patients by RF are the main goals of [32, 33],
respectively. El-Sappagh et al. [34] use RF in the simple form
for the detection of Alzheimer’s disease progression. In [35],
2.1. Astronomy, Bioinformatics, and Economics fields. In the RF is introduced as one useful machine learning tool for
astronomy field, Markel and Bayless [7] use RF for the healthcare domain, especially for COVID-19 modeling.
classification of type IA and core-collapse supernovae. Chen Khedkar et al. [36] use Patients Electronic Health Records
et al. [8] propose an approach to detect the potential signal for predicting the heart failure risks by RF. Hane et al. [37]
photons by RF. In the bioinformatics, Pang et al. [9] propose propose a model for prediction of the dissolution behaviour
a method to mitigate the computational complexity of RNA of a wide variety of oxide glasses. Subudhi et al. [38] propose
simulation software by a typical random forest. Darmawan a method by RF to detect the ischemic stroke by a sequence
et al. [10] propose an age estimation model in the bio- of MRI images. Javadi et al. [39] propose a method to predict
informatics field. In the economics field, Park et al. [11] the immunogenic peptides of intracellular parasites. Iden-
propose two stages of short-term load forecasting by random tifying the key risk factors associated with acute rejection in
forest and deep neural networks to reduce energy costs. 12 organ transplantation is the main propose of [40]. In Singh
use a typical RF to solve the e-commerce product classifi- et al. [41], RF has been used as one of the classifiers to classify
cation problem. Modeling consumer credit risk by RF is the the covid-19 spread. Na et al. [42] propose an automatic
main goal of [13]. 14 increase tree correlation by controlling walking mode change of the above-knee prosthesis. Clus-
the probability of placing splits along with strong predictors tering and predicting vital signs by RF is the goal of [43]. Zhu
to deal with high-dimensional settings. Sikdar et al. [15] et al. [44] optimize the parameters of the random forest by
proposed a variable selection method based on RF to identify improved fish Swarm algorithm for predicting the knee
the key predictors of price change in amazon. contact force. A method for identifying foreign particles for
quality detection of liquid pharmaceutical products is pre-
sented by [45]. Lee and Jung [46] consider the relation
2.2. General and Global Problem fields. In the general field,
between teacher attachment and student growth. 47 propose
Giffon et al. [16] use the mean of orthogonal matching
a practical method for SIF downscaling. Guanter et al. [48]
pursuit algorithms for calculating the weights of the linear
present a method based on RF for predicting diabetes. Subasi
combination for producing a linear combination of trees
et al. [49] propose a decision support system for the diag-
with minimum training error. Combining RF and gener-
nosis of migraine by RF. Classification of the driver’s stress
alized linear mixed models is the main idea of [17] to model
level is the main goal of [50]. Ayata et al. [51] propose an
clustered and longitudinal binary outcomes. Mohapatra
emotion recognition algorithm from multimodal physio-
et al. [18] optimize the random forest by use of unequal
logical signals by using the random forest as one of the
weight voting strategy. Ji et al. [19] propose a hybrid model
machine learning methods for recognition.
for crowd counting by a combination of convolutional
neural networks (CNN) and deep regression forest. Santra
et al. [20] propose a deterministic dropout to remove un- 2.4. Industrial and Network fields. Zeraatpisheh et al. [52]
important connections in NN by RF. Proposing the oblique use typical RF for producing the feature map in the in-
RF without explicit regularization techniques by minimizing dustrial field. Du et al. [53] propose a rapid and accurate
the structural risk is the main goal of [21]. Katuwal et al. [22] detection technique for pesticide detection by RF to
4 Computational Intelligence and Neuroscience

Let D � {(x1, y1), (x2, y2), . . ., (xN, yN)} denote the training data, with xi � (xi,1, xi,2, . . ., xi,p)T
For j � 1 to J:
Take a bootstrap sample D of size N from D.
Using the bootstrap sample, Dj as the training data fit a tree.
(a) Start with all observations in a single node.
(b) Repeat the following steps recursively for each node until the stopping criterion is met: (i) Select m predictors at random from
the p available predictors.
Find the best binary split among all binary splits in the predictors from step (i).
Split the node into two descendant nodes using the split from step (ii).
To make a prediction at a new point x.
􏽢
f(x) � argmaxy 􏽐Jj�1 I(h􏽢j (x))
Where h􏽢j (x) is the prediction of the response variable at x using the jth tree.

ALGORITHM 1: The random forest pseudocode for classification applications [1].

construct a quantitative detection model. Improving the classifiers is presented by [76] for finding the best classifiers in
performance of mapping for mineral is the main goal of the subject literature of text classification. The random forest
reference [54]. Liu et al. [55] propose an adaptive electrical is used as one of the base learners of the ensemble model for
period partition algorithm for open-circuit fault detection. fake news detection by [77]. Analyzing the reviewer’s
Software fault prediction by ensemble techniques is inves- comment for sentiment analysis is the main goal of [78].
tigated by [56]. In [57], the RF id is used to build a dis- Zhang et al. [79] propose two novel label flipping attacks to
tributed energy system. A comprehensive image processing evaluate the robustness of NB under noise by random forest.
model is proposed by [58]. Ho et al. [59] uses RF to propose a Recognizing newspaper text by RF is done by [80]. Mad-
framework that uses climate data to model hydropower ichetty and Sridevi [81] use RF as one of the classifiers for
generation. Zhou et al. [60] use RF for small and unbalanced detecting the damage assessment tweets. Madasu and Elango
datasets to create a risk prediction model for decision- [82] use the typical RF for feature selection for sentiment
making tool. Deng et al. [61] propose an authentication analysis. Chang et al. [83] use online customer reviews for
method for protecting high-value food products by RF. The opinion mining by RF. Text classification by simple RF is the
forecast for agricultural products by RF is proposed by [62]. goal of [84]. Onan and Toçouglu [85] present a method for
Jeong and Kim [63] use weighted random forest for the link document clustering and topic modeling on massive open
prediction model. Khorshidpour et al. [64] present an ap- online courses. Sentiment analysis of technical words in
proach to model an attack against classifiers with non- English by the Gini index for feature selection is done by [86].
differentiable decision boundary. Fusing multi-domain Beck [87] uses ensemble learning and deep learning for
entropy and RF is the main goal of [65] for proposing a fault sentiment classification scheme with high predictive per-
diagnosis method of the inter-shaft bearing. Analyzing the formance in massive open online courses’ reviews. Onan [88]
wine quality is presented by [66]. In the network field, present a deep learning based approach to sentiment analysis.
Madhumathi and Suresh [67] develop a model to predict the This approach uses TF-IDF weighted Glove word embedding
future location of a dynamic sensor node in wireless with CNN LSTM architecture. Onan and Tocoglu [89]
communications. Fang et al. [68] propose an encrypted present an effective sarcasm identification framework on
malicious traffic identification method. Detecting the in- social media data by pursuing the paradigms of neural
trusion in the network by typical RF is proposed by [69], and language models and deep neural networks. In the tourism
intrusion detection in the network security by tuning the RF field, Rodriguez-Pardo et al. [90] propose a method based on
parameter of the Moth-Flame optimization algorithm is simple RF for predicting the behaviour of tourists. Predicting
presented by [70]. the travel time to reduce traffic congestion is the main goal of
[91]. Jamatia et al.92 propose a method for tourist destina-
tions’ prediction. In urban planning, Baumeister et al.93 rank
2.5. Physics, Text Processing, Tourism, and Urban Planning the urban forest characteristics for cultural ecosystem ser-
fields. In the physics field, Mingjing [71] measure and vices supply by typical RF. Forecasting road traffic conditions
quantify the pH of soil by RF. 72 propose a model for in done by [94]. The simulation of urban space development
extracting complex relationships between energy modu- by RF is presented by [95]. Investigating the information on a
lation and device efficiency. Zhang et al. [73] propose a gross domestic product for the analysis of economic devel-
model to accurately and effectively predict the UCS of opment is presented by [96]. Mei et al. [97] propose a method
LWSCC by a beetle antennae search algorithm for tuning to identify the spatiotemporal commuting patterns of the
the hyper-parameters of RF. The prediction of geotechnical transportation system. In this brief review, the mentioned
parameters by typical RF is made by [74]. Creep index references are categorized in terms of innovation and
prediction by the RF algorithm to determine the optimal functionality.
combination of variables is the main goal of [75]. In the text As can be seen from Table 1, RF has a high range of
processing field, the comparison between RF and other applications and variations in scope. In contrast, both in
Computational Intelligence and Neuroscience 5

Table 1: Brief review of RF literature on functionality and innovation.

Type Field Paper
Astronomy [7], [8]
Bioinformatics [9], [10]
Economics [11], [12], [13]
Global problem [26], [27], [28]
[32], [33], [34], [35], [36], [41], [98], [37], [39], [40], [42], [43], [45], [46], [47],
Healthcare
[48], [49], [50], [51],
Functionality
Industrial [52], [53], [54], [55], [56], [57], [58], [59], [60], [61], [62]
Network [63], [67], [68], [69], [99], [100]
Physics [71], [72]
Text processing [76], [77], [78], 80 [81], [82], [83], [84]
Tourism [91], [92]
Urban planning [93], [94], [95], [96], [97]
Economics [14], [15]
General [16], [17], [18], [19], [101], [21], [22], [23], [24], [25]
Global problem [30], [31]
Healthcare [44]
Innovative method
Industrial [65]
Network [64]
Physics [73], [75]
Text processing [79], [86]

terms of quantity and quality, their innovations are often Environment’s

limited to set various parameters, and there is no significant response β
innovation in the base learner combinations.
Learning
Environment
automata

3. Learning Automata α
Learning automata’s
action
Learning Automata (LA) is one of the learning algorithms
that, after selecting different actions at different times, Figure 1: Interaction of learning automata with the environment.
identify the best practices in terms of responses received
from a random environment. LA selects an action from the
set of actions in the vector of probabilities, and this action is
is the probability vector of the LA actions and
evaluated in the environment. By using the received signal
from the environment, the LA updates the probability vector P(n + 1) � T[P(n), α(n), β(n)], (12)
and, by repeating this process, the optimal action is gradually
identified. The classification problem can be formulated as a
team of LA that operates collectively to optimize an objective is the learning algorithm.
function [102]. In Figure 1, the interaction of the learning In LA, three different models can be defined in the
automata and the environment is shown. environment. In the P-Model, the environment presents the
Finding the global optimum in the solution space is values of 0 or 1 as the output. In the Q-Model, the output
another advantage of using the LA. The LA can be formally values of the environment are discrete numbers between 0
represented by the quadruple and 1. In the S-Model, the output of the environment is the
continuous value between 0 and 1. The selected actions by
LA � 􏼈α, β, P, T􏼉, (8) the LA are updated by both the signal received from the
environment and using reward and penalty functions. The
in which amount of allocated reward and penalty to the LA action can
α � 􏼈α1 , α2 , . . . , αr 􏼉 (9) be defined in four ways: LRP, where the number of rewards
and penalties are considered the same; LRεP in which the
is the set of actions (outputs) of the LA; in other words, the amount of penalty is several times smaller than the reward;
set of inputs of the environment. LRI in which the penalty amount is considered 0; and LIP,
where the reward amount is considered 0 [103].
β � 􏼈β1 , β2 , . . . , βr 􏼉, (10)
At each instant n, the action probability vector pi(n) is
is the set of inputs of the LA; in other words, the set of updated by the linear learning algorithm given in equation
outputs of the environment. (13) if the chosen action ai(k) is rewarded by the environ-
ment, and it is updated according to equation (14) if the
P � 􏼈p1 , p2 , . . . , pr 􏼉, (11) chosen action is penalized [104].
6 Computational Intelligence and Neuroscience

base learners and randomly injected into the base learners.

⎨ pi (n + 1) � pi (n) + a􏼂1 − pi (n)􏼃,
⎧
⎪
(13) The difference between this step and the similar step in the
⎪
⎩ p (n + 1) � (1 − a)p (n), ∀ j, j ≠ i,
j j
random forest is that all the features in the feature vector are
given to all base learners, and the feature replacement option
⎧
⎪ pj (n + 1) � (1 − b)pi (n), is removed.
⎪
⎨
(14) After the first run, the prediction models are created in
⎪
⎩ pj (n + 1) � b +(1 − b)pj (n), ∀ j ; j ≠ i,
⎪ the base learners and placed in a pool that is actually an
r−1 interactive environment with the learning automata. The
, where “a” is the reward parameter, “b” is the penalty results obtained from the base learners for each new sample
parameter, and “r” is the number of actions. The authors are given in the form of a reinforcement signal to the
applied the LA in the proposed method, because: learning automata, which we know as the primary feedback
(i) The LA presents an acceptable performance in of the environment. Depending on whether the received
uncertain situations. reinforcement signal is a reward or a penalty, the chances of
selecting each of the base learners, -which they are the
(ii) The LA does search action in the probability space.
actions of the learning automata - are updated. It should be
(iii) The LA requires simple feedback from the envi- noted that the initial probability of selecting these actions is
ronment to optimize its state. considered equal at the start. If we have R base learners to
(iv) Since the LA has a simple structure, it has a simpler form the ensemble, the probability of the initial selection of
implementation in both software and hardware. each of them is equal to
(v) The LA is not constrained to use accuracy criteria p DTr 􏼁 � (1/R). (15)
for optimization usage.
(vi) The LA is applicable in real-time usage since the LA It is clear that the sum of the probabilities of all actions
is not involved with light computational complexity will be equal to 1.
[105]. R
􏽘 (pDT) � 1. (16)
i�1
4. Proposed Method
The initial probability of selecting actions is considered
The random forest is one of the methods of ensemble equal because all of them are homogeneous in terms of
learning that all constructor classifiers are same type (i.e., separating power.
decision tree). Therefore, the random forest is a homo- In the proposed method, integration of the base learners
geneous ensemble learning method. In this article, we is performed by the LA. Therefore, for each input in the test
intend to use the idea of reinforcement learning to in- set, a linear LA is defined, and the action of each LA cor-
crease the efficiency of random forest and add the ability responds to selecting the base learners. The process of
to adapt to the conditions of the problem for this data running base learners and receiving feedback from the
mining algorithm. The details of the proposed method are environment continues until the probability of selecting
described below. actions converges to one of the base learners, or the number
The method proposed in this paper is based on the idea of repetitions for learning automata exceeds the pre-
of reinforcement learning, and it employs the learning determined limit. Once the probability of selections con-
automata to implement the idea. The learning automata is verges, then the result of the base learner, which has the
the core of the proposed method, and by receiving feedback highest probability of selection, is determined as the result of
from the environment for each action, it updates the the ensemble for that particular input. In such a case, finding
probability selection of the actions. In the proposed method, the global optimal is guaranteed by the algorithm, and
each base learner, all of which are decision tree, are con- because all the features in the feature vector are examined,
sidered as learning automata actions. rare modes are also covered, and the ability to adapt to the
In the proposed method, the training data are first conditions of the problem space and independence from the
randomly divided into N sections. In this division, N cor- domain is stabilized. In the proposed method, the random
responds to the number of trees we want to have in the selection of subsets causes interdependence between trees.
forest. Unlike the random forest, in which the predictive The depth of all the decision trees in the proposed method is
model works by averaging or voting between trees, in the considered equal. Each decision tree divides the training
proposed method, the predictive model is created using data differently at the leaf level. The pseudocode of the
learning automata, which forms the core of the algorithm. proposed method is shown in Algorithm 2.
The block diagram of the proposed method is shown in In the learning automata block in Figure 2, there are two
Figure 2. functions called the reward function and penalty function.
The preprocessing step in the proposed method is a Activation of one of these two functions is based on the type
general step, and based on what type of data the processing of reinforcement signal received from the environment. The
area is dealing with, the details of this phase are determined. received signal from the environment determines whether
In the proposed method, at first, similar to the random forest the result of the base learner activity or the selected action in
method, the training data are divided into the number of the previous step was useful or not. If the result is useful, that
Computational Intelligence and Neuroscience 7

Splitting

D2
D
Dt D1 D
Training DTr D4 Preprocessing Data
Di
Trained D3 D5
DTr
Learning automata

Updating
DT1
Penalty Reward
Environment (pool)

DT3
DT2 function function
β
DT5
DT4 α1 α2 αr
α …
DTi

No
Convergence?

Yes

Result
Figure 2: The block diagram of the proposed method.

Input D = {(x1, y1), (x2, y2), . . ., (xN, yN),} denote the training data with xi � (xi,1, xi,2, . . ., xi,p)
(1) Output classified test data
(2) Assumption
(3) LA : Learning automata
(4) DTr � {DT1, DT2, . . . , DTR} denote the base learners
(5) αi: LA action//Choose DTr
(6) a: Reward parameter
(7) b: Penalty parameter
(8) Pool : All the trained base learners
(9) Algorithm
(10) For r � 1 to R do
(11) Create a dataset Dt, by sampling (N/R) items, randomly with replacement from D
(12) Train DTr using Dt, and add to the pool
(13) end//for
(14) For each test sample
(15) {
(16) LA � new LA//Create an LA object from LA class
(17) While ((LA convergences to an action) or (LA exceeds predefined iteration number))
(18) {
(19) Select one of the actions at random and execute it, by the LA, Let it be αi
(20) If (αi predicts the new test sample correctly) then//Update the probability of selection vector
(21) p (n + 1) � p (n) + a[1 − p (n)]
􏼨 pi (n + 1) � (1i − a)p (n), i ∀ j, j ≠ i //reward the selected αi
j j
(22)
else
(23) p (n + 1) � (1 − b)p (n),
􏼨 pj (n + 1) � (b/R − 1)i + (1 − b)p (n), ∀ j, j ≠ i, //Penalty the selected αi
j j
(24)
}//end while
(25) }//end for
(26) Return DTr
(27) Classified test data � the prediction of DTr
(28) End.//algorithm

ALGORITHM 2: The pseudocode of the proposed method.

8 Computational Intelligence and Neuroscience

action must be rewarded or, in other words, increase the considered to be zero, and the results of the proposed
probability of its selection. The increase in the probability of method in this mode are shown in Figure 3.
the selected action is determined by the parameters “a” and Based on the literature on learning automata in the LRεP
“b,” which are called the reward parameter and the penalty mode, the value of the penalty parameter is considered to be
parameter, respectively. much smaller than the value of the reward parameter. The
To comply with (16), that is, the sum of the probabilities results of the proposed method are shown in the LRεP mode
of all actions being equal to one, the probability of all other in Figure 4.
actions is reduced according to the size of the parameter “a.” As mentioned in the learning automata section, in the
If the result of the selected action is not useful, that action LRP mode, the values of the penalty and reward parameters
must also be penalized. In other words, the probability of are considered equal. The results of the proposed method in
that action must be reduced. To do this, the probability of this mode are also shown in Figure 5.
selecting that action is reduced to the size of parameter “b,” A comparison of the results obtained from the
and as a rewarding mode, and to observe (16), the probability implementation of the proposed method in three adjust-
of selecting other actions is increased by the size of the able modes for learning automata shows that the settings
parameter “b.” on the LRP mode have resulted in the highest accuracy for
In the proposed method, the learning automata model identification. Then there are LRεP and LRI modes. In the
environment is assumed to be the P-Model, where the LRεP mode, the setting a � 0.01, b � 0.01 is not considered,
environment defines zero and one values as outputs. Zero because these values are equal to the first values set in the
means reward, and one means penalty. If the correct answer LRP mode, and in order to prevent duplication of results in
is received from the selected base learner by the LA, the different tables, these settings have been removed from the
action of choice will be rewarded; otherwise, it will be LRεP mode. For this reason, the number of experiments
penalized. performed on LRεP mode evaluations is one less than the
other two. Considering that the settings of reward and
5. Evaluation penalty parameters in the LRP mode with the values of
a � 0.5, b � 0.5 have resulted in the highest efficiency,
In order to thoroughly evaluate the efficiency of the pro- evaluation has been done on other datasets with these
posed method, in this section, the details of the evaluation of settings. A comparison of the proposed method and
the proposed method are presented separately from the data similar approaches in the subject literature is shown in
used and the experimental results. Table 3.
As can be seen in Table 3 from the point of view of
5.1. Datasets. In order to evaluate the proposed method and accuracy, the proposed method offers better performance
to create an environment with the dynamic behaviour of than the methods available in the subject literature, which
data, different domains of applications have been selected. indicates an improvement in the aggregation model of the
As mentioned in the previous sections, dynamic behaviour base learners. This improvement is due to the use of rein-
refers to the different results that an instance exhibits in forcement learning ideas of the method of aggregation of
different environmental conditions. Variety in the results of basic classifiers, which is known as base learner. The use of
different environments is created by a specific domain. Text reinforcement learning ideas has improved the ability of the
data are one of the most well-known types of data that created ensemble, and it improved the ability to address
exhibit such dynamic behaviour. In other words, these types issues in which data exhibit dynamic behaviour. The results
of data are one of the optimal options for creating a dynamic of experiments performed on different data confirm the
environment, which proves the adaptability of the proposed capabilities added to the random forest by the proposed
method. The details of the selected data for the evaluation method. As mentioned earlier, in the field of opinion
phase are shown in Table 2. mining, the type of text data is the most obvious data that
exhibit such dynamic behaviour. Therefore, the optimal
5.2. Experimental Result. In order to evaluate the proposed values for the reward and penalty parameters have been
method, eighteen datasets in different domains introduced determined in these types of data, and these settings have
in the previous section have been used. In the literature on been used for other types of data.
learning automata, different modes have been considered for In addition to the accuracy criterion, other statistical
tuning learning automata; in this paper, three modes have criteria have been examined to evaluate the proposed
been used to evaluate the proposed method. The LIP mode is method. As can be seen in Table 4, the proposed method has
not considered due to poor results. The evaluation results of shown better results in both positive and negative classes
each of the LRI, LRεP, and LRP modes are shown in separate than the methods available in the literature. Among the
figures. In order to determine the optimal value for the statistical criteria, Precision (P) determines the exactness of
reward and penalty parameters, six text datasets have been the results obtained from the classifier, and Recall (R) de-
selected. The reason for this choice is the high diversity in the termines the completeness of the results obtained from the
behaviour of textual data as well as a large number of classifier. The results obtained from the test in the mentioned
samples and a large number of features of these six datasets. statistical criteria show that the proposed method has a high
In the LRI mode, the value of the penalty parameter is performance.
Computational Intelligence and Neuroscience 9

Table 2: Details of textual data used for evaluation.

Domain Name # Feature # Instance
Stanford—Sentiment 140 corpus [106] Bag of word 1600000
Large dataset of movie reviews [107] Bag of word 50000
Sentence polarity dataset v1.0 [108] Bag of word 10662
Text
Internet movie database [105] Bag of word 1400
Yelp review [105] Bag of word 598000
Amazon review [105] Bag of word 1000000
Heart disease dataset [105] 13 200
Breast cancer dataset [105] 30 569
Arrhythmia dataset [105] 279 454
Parkinson dataset [105] 45 241
Healthcare
Caesarean section dataset [105] 5 81
Gene expression dataset [105] 255 801
Diabetes dataset [105] 7 765
Statlog (heart) dataset [105] 13 271
Ionosphere dataset [105] 34 352
Physical
Sonar, mines vs. rocks dataset [105] 60 208
Voice dataset [105] 20 3168
Sound
Emotions from music dataset [105] 28 592

100

90

80

70

60
Accuracy

50

40

30

20

10

0
a = 0.01, b = 0 a = 0.05, b = 0 a = 0.1, b = 0 a = 0.3, b = 0 a = 0.5, b = 0 a = 0.7, b = 0
Sentiment 140 dataset 74.05 74.05 74.05 74.05 74.05 74.05
Large dataset of movie reviews 86.13 85.98 85.98 85.98 85.98 85.98
Sentence polarity dataset 72.94 72.94 72.94 72.94 72.94 72.94
Movie reviews dataset 82.31 82.31 82.31 82.31 82.31 82.31
Yelp review polarity 89.56 89.56 89.56 89.56 89.56 89.56
Amazon review polarity 81.41 81.41 81.41 81.41 81.41 81.41

Figure 3: The results of the proposed method in LRI mode.

6. Discussion ranking the set of these parameters based on their

performance.
In this section, more details of the proposed method are
explained along with the reasons for the need to address
these details. These include the details of the preprocessing 6.1. Preprocessing. As explained in the proposed method
step, tuning the learning automata parameters, as well as section, the preprocessing step is a general step. In order for
10 Computational Intelligence and Neuroscience

100
90
80
70
60

Accuracy
50
40
30
20
10
0
a = 0.05, b = 0.01 a = 0.1, b = 0.01 a = 0.3, b = 0.01 a = 0.5, b = 0.01 a = 0.7, b = 0.01

Sentiment 140 dataset 74.4 74.65 74.9 74.55 73.95

Large dataset of movie reviews 86.57 86.03 86.33 86.47 86.57
Sentence polarity dataset 73.44 74.53 73.74 73.29 75.53
Movie reviews dataset 82.31 82.67 81.22 82.31 81.58
Yelp review polarity 90.32 89.45 89.23 89.78 89.23
Amazon review polarity 81.33 81.75 80.83 80.5 80.83

Figure 4: The results of the proposed method in the LRεP mode.

100
90
80
70
60
Accuracy

50
40
30
20
10
0
a = 0.01, a = 0.05, a = 0.1, a = 0.3, a = 0.5, a = 0.7,
b = 0.01 b = 0.05 b = 0.1 b = 0.3 b = 0.5 b = 0.7
Sentiment 140 dataset 74.7 74.8 75.85 76.85 76.3 75.65
Large dataset of movie reviews 85.98 86.23 87.06 87.35 86.62 86.82
Sentence polarity dataset 74.33 75.53 75.83 76.48 77.03 76.08
Movie reviews dataset 81.94 81.58 83.75 83.03 85.92 83.75
Yelp review polarity 89.78 89.67 89.34 89.78 90.76 89.67
Amazon review polarity 80.41 80.58 81.08 81.16 82.58 81.33

Figure 5: The results of the proposed method in the LRP mode.

the evaluation, different data from different domains were Repository [109]. And their basis for accuracy is based on
examined. The preprocessing of textual data, along with the previous research works that have used these data.
relevant details, is described below. It should be noted that In order to prepare textual data for the main process,
preprocessing for other types of data, such as feature ex- the opinion mining domain is selected and the related
traction, feature selection, normalization, noise removal, preprocessing is as follows. The details of the pre-
and other related preprocessing, has not been performed processing step for text data in opinion mining are shown
because all of them are taken as clean data from the UCI in Figure 6.
Computational Intelligence and Neuroscience 11

Table 3: Comparison of the proposed method with similar approaches in the subject literature.
Dataset Averaging Majority Voting Random Forest Our Method
Sentiment140 dataset 74.54 75.50 74.30 76.30
Large dataset of movie reviews 86.28 86.86 86.42 86.62
Sentence polarity dataset 73.75 74.63 73.38 77.03
Text
Movie reviews dataset 81.58 81.58 81.67 85.92
Yelp review polarity 89.47 90.32 89.74 90.76
Amazon review polarity 80.86 81.66 80.97 82.58
Heart disease dataset 58.00 57.50 57.50 65.00
Breast cancer dataset 97.41 97.36 96.49 98.24
Arrhythmia dataset 80.71 85.71 81.31 85.71
Parkinson dataset 63.95 64.58 64.58 68.75
Healthcare
Caesarean section dataset 60.31 62.50 43.75 68.75
Gene expression dataset 95.59 95.62 96.27 98.75
Diabetes dataset 75.77 75.32 74.67 76.62
Statlog (heart) data set 81.20 81.48 79.62 85.18
Ionosphere dataset 91.05 91.54 92.95 95.77
Physical
Sonar, mines vs. rocks dataset 85.23 85.71 73.80 88.09
Voice dataset 76.38 76.18 76.49 88.95
Sound
Emotions from music dataset 78.23 78.15 82.35 84.03

Table 4: Comparison of statistical criteria.

Positive class Negative class Positive class Negative class
Method P (%) R (%) F1 (%) P (%) R (%) F1 (%) Method P (%) R (%) F1 (%) P (%) R (%) F1 (%)
Sentiment140 dataset Parkinson dataset
MV 72.44 70.35 71.38 70.90 72.96 71.92 MV 69.23 66.67 67.92 59.09 61.9 60.47
RF 72.44 74.58 73.49 76.46 74.43 75.43 RF 69.23 66.67 67.92 59.09 61.9 60.47
OM 75.20 80.36 77.69 81.17 76.16 78.59 OM 76.92 68.97 72.73 59.09 68.42 63.41
Large dataset of movie reviews Caesarean section data set
MV 87.50 87.33 87.41 87.41 87.59 87.50 MV 55.56 71.43 62.5 71.43 55.56 62.5
RF 85.83 76.16 80.70 73.37 83.93 78.29 RF 33.33 50 40 57.14 40 47.06
OM 87.80 87.97 87.88 88.10 87.93 88.01 OM 66.67 75 70.59 71.43 62.5 66.67
Sentence polarity dataset Gene expression dataset
MV 75.63 72.83 74.20 72.80 75.61 74.18 MV 92.31 94.12 93.2 97.25 96.36 96.8
RF 74.62 67.78 71.08 65.95 72.94 69.27 RF 94.23 93.23 94.23 97.25 97.25 97.25
OM 76.04 73.29 74.64 73.29 76.04 76.64 OM 98.08 98.08 98.08 99.08 99.08 99.08
Movie reviews dataset Diabetes dataset
MV 83.10 82.52 82.81 81.84 82.09 81.78 MV 90.29 76.86 83.04 45.1 69.7 54.76
RF 77.46 71.43 74.32 76.41 73.98 70.54 RF 90.29 76.23 82.67 43.14 68.75 53.01
OM 84.51 83.33 83.92 82.22 83.46 82.84 OM 91.26 77.69 83.93 47.07 72.73 57.14
Yelp review polarity Voice dataset
MV 90.85 88.04 88.42 87.11 90.11 88.59 MV 64.58 76.09 69.86 84.85 76.24 80.31
RF 80.64 85.94 83.21 86.22 81.00 83.53 RF 64.58 76.75 70.14 85.4 76.35 80.62
OM 90.00 89.43 89.71 88.89 89.49 89.19 OM 87.82 86.55 87.18 89.81 90.81 90.3
Amazon review polarity Emotions from music dataset
MV 82.68 79.45 81.03 79.38 82.62 80.97 MV 83.58 78.87 81.16 71.15 77.08 74
RF 79.12 74.56 76.77 73.98 87.61 76.22 RF 88.06 81.94 84.89 75 82.98 78.79
OM 83.70 79.52 81.56 79.21 83.45 81.28 OM 92.54 81.58 86.71 73.08 88.37 80
Heart disease dataset Sonar, mines vs. Rocks dataset
MV 61.90 59.09 60.47 52.63 55.56 54.05 MV 87.5 87.5 87.5 83.33 83.33 83.33
RF 61.90 59.09 60.47 52.63 55.56 54.05 RF 79.17 76 77.55 66.67 70.59 68.57
OM 61.90 68.42 65.00 68.42 61.90 65.00 OM 87.5 87.5 87.5 83.33 83.33 83.33
Breast cancer data set Statlog (heart) data set
MV 95.74 97.83 96.77 98.51 97.06 97.78 MV 79.19 79.19 79.19 83.33 83.33 83.33
RF 97.87 93.88 95.83 95.52 98.46 96.97 RF 75 78.26 76.6 83.33 80.65 81.67
OM 97.87 97.87 97.87 98.51 98.51 98.51 OM 79.19 86.36 82.61 90 84.38 87.1
Arrhythmia data set Ionosphere data set
MV 88.37 82.61 85.39 83.33 88.89 86.02 MV 82.14 95.83 88.46 97.67 89.36 93.33
RF 83.72 78.26 80.9 79.17 79.74 73.85 RF 85.71 96 90.57 97.67 91.3 94.38
OM 93.02 80 86.02 79.19 92.68 85.39 OM 89.29 100 94.34 100 93.48 96.63
P, R, and F1 refer to Precision, Recall, and F1-score. MV: majority voting, RF: random forest, and OM: our method.
12 Computational Intelligence and Neuroscience

Expressive Preprocessing Lemmatization

lengthening

Emoticons Word net

Emoticons dictionary dictionary Stemming
handling

HTML markups Slangs Stem list Stopwords

removal dictionary removal

Punctuation
Slangs handling
handling

Figure 6: Details of the preprocessing step for text data.

Expressive Lengthening. Word lengthening or word reinforcement learning is implemented using learning
stretching refers to the words that are elongated to automata, all three adjustable modes of the parameters of
express a particular emotion strongly, and the words reward and penalty are examined. The results of these three
with wrong spellings are corrected and replaced with modes were presented in the experimental result section. In
their original words. this paper, Friedman test statistical verification is used to
Emoticons Handling. It refers to the emoticons men- determine which mode and which settings are best ad-
tioned in the text that are replaced with their meaning, justable for the reward and penalty parameters. The values
which makes it easier to analyze the emoticons. set for parameters “a” and “b” are shown in Table 5. De-
termining the numerical value of these parameters is based
HTML Markups Removal. HTML markups presented
on the subject literature of learning automata. Of course, a
in the text are removed as they do not have any sen-
wide variety of values can be considered for these two pa-
timental value attached to it.
rameters. In this paper, an attempt has been made to tune the
Slangs Handling. The slangs are used for writing a given parameters in such a way that all the modes are considered
word, in short syllables, which depict the same meaning so that they can be used to prove the efficiency of the
but save the time of typing. In slangs handling, the proposed method compared to the previous methods.
slangs presented in the text are replaced with their
original words.
Punctuation Handling. Punctuations are used in a text 6.3. Ranking. Friedman test statistical verification [110] is a
to separate sentences and their elements, and to clarify ranking method that, the difference between the ranks
their meaning. At punctuation handling, once the assigned to each of the input samples, determines the op-
apostrophes are handled, all the remaining punctua- timal level of each option. In this paper, this verification
tions and numbers are removed. method has been used to determine the optimal value of
Stopwords Removal. Stopwords do not carry much reward and penalty parameters as well as to compare the
meaning and have no importance in the text. Stop- proposed method with the conventional methods in the
words are removed to get a simplified text. subject literature of ensemble learning. The results are shown
in Table 6.
Stemming. It refers to finding out the root or stem of a As can be seen in Table 6, there is a significant difference
word. Removing various suffixes to reduce the number between the rankings of the proposed method and the
of words is the purpose of stemming. rankings of the traditional methods, which indicate an
Lemmatization. It returns the base or dictionary form improvement in the efficiency of the proposed method
of a word, which is known as the lemma. It is very compared to other methods. Among the three modes
similar to stemming, but it is more akin to synonym considered for tuning reward and penalty parameters, it is
replacement. observed that the rankings have increased in LRI, LReP, and
BoW creation. The bag of word creation is the latest LRP modes, respectively. In the LRP mode, where the values
preprocess that is performed on the text preparation. of the reward and penalty parameters are considered the
same, the highest efficiency is also observed. There is a
significant difference between the Mean Rank of the best set
6.2. Tuning the Parameters of Reward and Penalty. In the of the reward and penalty parameters in the proposed
subject literature of the learning automata, three different method and this rank in the random forest method. The
modes have been defined to tune the parameters of reward difference between the ranks is proof that the proposed
and penalty. In the proposed method, in which the idea of method is optimal versus the traditional methods of
Computational Intelligence and Neuroscience 13

Table 5: Numerical values tuned for reward and penalty parameters.

Mode Parameter
a 0 0.1 0.1 0.3 0.5 0.7
LRI
b 0 0 0 0 0 0
a 0.1 0.1 0.3 0.5 0.7
LRεP
b 0 0 0 0 0
a 0 0.1 0.1 0.3 0.5 0.7
LRP
b 0 0.1 0.1 0.3 0.5 0.7

Table 6: Friedman test statistical verification results for ranking the parameters of reward and penalty and comparing the proposed method
with the literature.
Method Tuning Mean rank Final rank
LRP a � 0.5, b � 0.5 19.17 1
LRP a � 0.3, b � 0.3 16.83 2
LRP a � 0.7, b � 0.7 15.58 3
MV Majority voting 14.67 4
LRP a � 0.1, b � 0.1 13.92 5
LReP a � 0.05, b � 0.01 12.17 6
LReP a � 0.1, b � 0.01 11.83 7
LReP a � 0.5, b � 0.01 10.08 8
LRP a � 0.05, b � 0.05 9.58 9
RF Random forest 9.17 10
LRP a � 0.01, b � 0.01 8.75 11
LIR a � 0.01, b � 0 8.42 12
LIR a � 0.05, b � 0 7.67 13
LIR a � 0.1, b � 0 7.67 13
LIR a � 0.3, b � 0 7.67 13
LIR a � 0.5, b � 0 7.67 13
LIR a � 0.7, b � 0 7.67 13
AV Averaging 7.58 14
LReP a � 0.3, b � 0.01 7.17 15
LReP a � 0.7, b � 0.01 6.75 16

Convergence rate Convergence rate

1.0 1.0

0.8 0.8
LA actions

LA actions

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
0 200 400 600 800 1000 0 200 400 600 800 1000
Number of iteration Number of iteration
(a) (b)
Figure 7: Continued.
14 Computational Intelligence and Neuroscience

Convergence rate Convergence rate

1.0 1.0

0.8 0.8
LA actions

LA actions
0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
0 200 400 600 800 1000 0 200 400 600 800 1000
Number of iteration Number of iteration
(c) (d)
Convergence rate Convergence rate

0.30
0.4
0.25
0.3 0.20
LA actions

LA actions 0.15
0.2
0.10
0.1
0.05
0.0 0.00
0 200 400 600 800 1000 0 200 400 600 800 1000
Number of iteration Number of iteration
(e) (f )
Convergence rate
1.0

0.8
LA actions

0.6

0.4

0.2

0.0
0 200 400 600 800 1000
Number of iteration
(g)

Figure 7: Convergence rate for different reward and penalty parameters. (a) a � 0.5, b � 0.5; (b) a � 0.3, b � 0.3; (c) a � 0.7, b � 0; (d) a � 0.1,
b � 0.1; (e) a � 0.01, b � 0; (f ) a � 0.05, b � 0.05; (g) a � 0.3, b � 0.

aggregating classifiers to achieve a strong classification number of iterations. As shown in Table 5, convergence at a
method. lower rate occurred in some of the other settings that scored
lower on the Friedman test.

6.4. Checking Convergence Rate. To more accurately address

the proposed method in terms of efficiency, LA convergence 6.5. Noise Resistance. In order to more accurately evaluate
has been investigated. Figure 7 shows the convergence of LA the proposed method and determine the resistance of the
actions for different amounts of reward and penalty vari- proposed method to noise, another evaluation has been
ables. In most of the different settings for these two pa- performed on the data presented in the previous section.
rameters, the convergence rate is high, and convergence to This evaluation was performed by injecting 20% noise into
one of the actions usually occurs before reaching a certain clean data. The results of the evaluation on noisy data show
Computational Intelligence and Neuroscience 15

100.00
90.00
80.00
70.00
60.00
Accuracy

50.00
40.00
30.00
20.00
10.00
-
Sentiment 140 dataset

Gene expression dataset

Diabete dataset
Sentence polarity dataset

Breast cancer dataset

Sonor, mines vs. Rocks dataset

Large dataset of movie reviews

Movie reviews dataset

Heart disease dataset

Parkinson dataset

Caesarian section dataset

Statlog (heart) dataset

Voice dataset

Emotions from music dataset

Ionosphere dataset
Arrhythmia dataset
Amazon review polarity
Yelp review polarity

Averaging Random forest

Majority voting Our method
Figure 8: The evaluation of the proposed method in the presence of noise.

that the proposed method, due to the use of learning learner. The choice of action is based on receiving feedback
automata, has high adaptability to the problem conditions, from the environment. This causes the dynamic behaviour of
and in the presence of noise, contrary to conventional data to be covered by using the idea of reinforcement
methods in the literature, the proposed method does not learning. On the other hand, given that at each stage,
suffer a sharp decline, and in such conditions, it shows high learning automata strives to achieve the highest amount of
efficiency compared to traditional methods. The evaluation achievable rewards, it is guaranteed to find the global optima
of the proposed method in the presence of noise is shown in in the problem space. Adaptability is another advantage of
Figure 8. the proposed method compared to similar methods in the
subject literature.
7. Conclusion and Future Work Due to the fact that in each step learning automata
operates based on environmental conditions and received
Base learner aggregation in ensemble learning should be feedback from the environment, the ability to adapt to the
done in such a way that the following points are met. First problem is met. The results of the evaluations performed in
point: selecting a base learner leads to the highest perfor- different data show that the proposed method has the ability
mance achievable in the current situation. Second point: if to achieve all the desired items mentioned above. Despite the
the situation changes due to the dynamics of the problem, fact that, unlike the random forest mechanism, all features are
the structure of the ensemble will change in such a way that it injected into all base learners in the proposed method, the
has the greatest amount of compatibility with the conditions efficiency of the proposed method in dealing with large-
of the new environment. Therefore, in order to meet the volume data has not decreased, and the results are more
above points and achieve an ensemble that is able to adapt to favorable than the classical methods. The proposed method is
the dynamic conditions of the problem, in this paper, a new independent of the data type and has the ability to handle any
method based on the idea of reinforcement learning is other type of data in any field. In order to substantiate this
proposed to integrate the base learners in the random forest. claim, and in order to evaluate the proposed method, different
In the proposed method, learning automata is used to re- types of data have been chosen. However, there are no re-
ceive feedback from the environment and perform actions strictions on the proposed method for dealing with different
on it. The general procedure is to receive feedback from the types of data. In this paper, a new method for aggregating the
environment, where the environment is a set of base learners base learners of the random forest using learning automata is
that we intend to combine to achieve a better performance proposed. Determining the optimal value for the parameters
than individual base learners. Learning automata actions of reward and penalty in the form of self-tuning is one of the
include choosing one of the base learners as the best base future works that the authors intend to do.
16 Computational Intelligence and Neuroscience

Data Availability [15] S. Sikdar, V. Kadiyali, and G. Hooker, Price Dynamics on

Amazon Marketplace: A Multivariate Random Forest Vari-
The authors declare that all the data are available publicly at able Selection Approach, 2019.
the UCI repository. [16] L. Giffon, C. Lamothe, L. Bouscarrat, P. Milanesi,
F. Cherfaoui, and S. Koço, Pruning Random Forest with
Conflicts of Interest Orthogonal Matching Trees, 2020.
[17] J. L. Speiser, B. J. Wolf, D. Chung, C. J. Karvellas, D. G. Koch,
The authors declare that there are no conflicts of interest and V. L. Durkalski, “BiMM forest: a random forest method
regarding the publication of this paper. for modeling clustered and longitudinal binary outcomes,”
Chemometrics and Intelligent Laboratory Systems, vol. 185,
pp. 122–134, 2019.
References [18] N. Mohapatra, K. Shreya, and A. Chinmay, “Optimization of
[1] C. Zhang and Y. Ma, Ensemble Machine Learning: Methods the random forest algorithm,” in Advances In Data Science
and Applications, Springer, New York, NY, USA, 2012. And Management, pp. 201–208, Springer, New York, NY,
[2] A. Galicia, R. Talavera-Llames, A. Troncoso, I. Koprinska, USA, 2020.
and F. Martı́nez-Álvarez, “Multi-step forecasting for big data [19] Q. Ji, T. Zhu, and D. Bao, “A hybrid model of convolutional
time series based on ensemble learning,” Knowledge-Based neural networks and deep regression forests for crowd
Systems, vol. 163, pp. 830–841, 2019. counting,” Applied Intelligence, pp. 1–15, 2020.
[3] M. Savargiv, B. Masoumi, and M. R. Keyvanpour, “A new [20] B. Santra, A. Paul, and D. P. Mukherjee, “Deterministic
ensemble learning method based on learning automata,” dropout for deep neural networks using composite random
Journal of Ambient Intelligence and Humanized Computing, forest,” Pattern Recognition Letters, vol. 131, pp. 205–212,
pp. 1–16, 2020. 2020.
[4] M. Kang, J. Ahn, and K. Lee, “Opinion mining using en- [21] M. A. Ganaie, M. Tanveer, and P. N. Suganthan, “Oblique
semble text hidden Markov models for text classification,” decision tree ensemble via twin bounded SVM,” Expert
Expert Systems with Applications, vol. 94, pp. 218–227, 2018. Systems with Applications, vol. 143, Article ID 113072, 2020.
[5] Y. Zhang, D. Miao, J. Wang, and Z. Zhang, “A cost-sensitive [22] R. Katuwal, P. N. Suganthan, and L. Zhang, “Heterogeneous
three-way combination technique for ensemble learning in oblique random forest,” Pattern Recognition, vol. 99, Article
sentiment classification,” International Journal of Approxi- ID 107078, 2020.
mate Reasoning, vol. 105, pp. 85–97, 2019. [23] P. Probst, M. N. Wright, and A.-L. Boulesteix, “Hyper-
[6] T.-H. Lee, A. Ullah, and R. Wang, “Bootstrap Aggregating parameters and tuning strategies for random forest,” Wiley
and Random Forest, pp. 389–429, Springer, New York, NY, Interdisciplinary Reviews: Data Mining and Knowledge,
USA, 2020. vol. 9, no. 3, Article ID e1301, 2019.
[7] J. Markel and A. J. Bayless, “Performance of random forest [24] S. Kim, M. Jeong, and B. C. Ko, “Interpretation and sim-
machine learning algorithms in binary supernovae classifi- plification of deep forest,” 2020, http://arxiv.org/abs/2001.
cation,” 2019, http://arxiv.org/abs/1907.00088. 04721.
[8] B. ChenZ. Li et al., “Forest signal detection for photon [25] V. Jain, J. Sharma, K. Singhal, and A. Phophalia, “Expo-
counting LiDAR using Random Forest,” Remote Sensing nentially weighted random forest,” Pattern Recognition And
Letters, vol. 11, no. 1, pp. 37–46, 2020. Machine Intelligence, pp. 170–178, Lecture Notes in Com-
[9] W. Pang, X. Liu, Z. Wang, Y. Fan, and J. Wang, “Predicting puter Science, vol. 11941, Springer, Cham, Switzerland, 2019.
RNA molecular specific hybridization via random forest,” in [26] M. StafoggiaP. Glantz et al., “A random forest approach to
Proceedings of the 2019 IEEE 7th International Conference on estimate daily particulate matter, nitrogen dioxide, and
Bioinformatics and Computational Biology (ICBCB),
ozone at fine spatial resolution in Sweden,” Atmosphere,
pp. 35–38, Hangzhou, China, 2019.
vol. 11, no. 3, p. 239, 2020.
[10] M. F. Darmawan, A. F. Zainal Abidin, S. Kasim, T. Sutikno,
[27] S. Hauglin and P. Montesano, “Modelling above-ground
and R. Budiarto, “Random forest age estimation model based
biomass stock over Norway using national forest inventory
on length of left hand bone for Asian population,” Inter-
data with ArcticDEM and Sentinel-2 data,” Remote Sensing of
national Journal of Electrical and Computer Engineering
(IJECE), vol. 10, no. 1, p. 549, 2020. Environment, vol. 236, p. 111501, 2020.
[11] S. Park, J. Moon, S. Jung, S. Rho, S. W. Baik, and E. Hwang, [28] H. R. Breidenbach and M. M. Saravi, “Land-subsidence
“A two-stage industrial load forecasting scheme for day- spatial modeling using the random forest data-mining
ahead combined cooling, heating and power scheduling,” technique,” Spatial Modeling In GIS and R for Earth And
Energies, vol. 13, no. 2, p. 443, 2020. Environmental Sciences, pp. 147–159, Elsevier, Amsterdam,
[12] J. Dai, T. Wang, and S. Wang, “A deep forest method for Netherlands, 2019.
classifying e-commerce products by using title information,” [29] J. Cai, K. Xu, Y. Zhu, F. Hu, and L. Li, “Prediction and
in Proceedings of the 2020 International Conference On analysis of net ecosystem carbon exchange based on gradient
Computing, Networking And Communications (ICNC), boosting regression and random forest,” Applied Energy,
pp. 1–5, Big Island, HI, USA, 2020. vol. 262, Article ID 114566, 2020.
[13] M. Papoušková and P. Hajek, “Modelling loss given default [30] L. Wen and X. Yuan, “Forecasting CO2 emissions in Chinas
in peer-to-peer lending using random forests,” in Proceed- commercial department, through BP neural network based
ings of the Intelligent Decision Technologies 2019, pp. 133–141, on random forest and PSO,” Science of The Total Environ-
Springer, Malta, Europe, 2020. ment, vol. 718, Article ID 137194, 2020.
[14] D. Borup, B. J. Christensen, N. Mühlbach, and M. S. Nielsen, [31] Y.-S. Li, H. Chi, X.-Y. Shao, M.-L. Qi, and B.-G. Xu, “A novel
The Effects Of Targeting Predictors In A Random Forest random forest approach for imbalance problem in crime
Regression Model, 2020. linkage,” Knowledge-Based System, Article ID 105738, 2020.
Computational Intelligence and Neuroscience 17

[32] S. K. Mohapatra and M. N. Mohanty, “Big data analysis and [47] X. LiuL. Liu et al., “Downscaling of solar-induced chloro-
classification of biomedical signal using random forest al- phyll fluorescence from canopy level to photosystem level
gorithm,” New Paradigm In Decision Science And using a random forest model,” Remote Sensing of Environ-
Management, pp. 217–224, Springer, New York, NY, USA, ment, vol. 231, Article ID 110772, 2019.
2020. [48] S. Guanter and J. Santosh Kumar, “Performance evaluation
[33] A. Joshi, T. Choudhury, A. Sai Sabitha, and K. Srujan Raju, of random forest with feature selection methods in pre-
“Data mining in healthcare and predicting obesity,” in diction of diabetes,” International Journal of Electrical and
Proceedings of the Third International Conference on Com- Computer Engineering, vol. 10, 2020.
putational Intelligence and Informatics, pp. 877–888, [49] A. Subasi, A. Ahmed, E. Aličković, and A. Rashik Hassan,
Hyderabad, India, 2020. “Effect of photic stimulation for migraine detection using
[34] S. El-SappaghR. Sahal et al., “Alzheimer’s disease progression random forest and discrete wavelet transform,” Biomedical
detection model based on an early fusion of cost-effective Signal Processing and Control, vol. 49, pp. 231–239, 2019.
multimodal data,” Future Generation Computer Systems, [50] N. El Haouij, J.-M. Poggi, R. Ghozi, S. Sevestre-Ghalila, and
vol. 115, pp. 680–699, 2021. M. Jaı̈dane, “Random forest-based approach for physio-
[35] Y. Saleh, A. Halidou, and P. T. Kapen, “A review of logical functional variable selection for driver’s stress level
mathematical modeling, artificial intelligence and datasets classification,” Statistical Methods & Applications, vol. 28,
used in the study, prediction and management of COVID- no. 1, pp. 157–185, 2019.
19,” Applied Intelligence, vol. 50, no. 11, pp. 3913–3925, 2020. [51] D. Ayata, Y. Yaslan, and M. E. Kamasak, “Emotion recog-
[36] S. Khedkar, P. Gandhi, G. Shinde, and V. Subramanian, Deep nition from multimodal physiological signals for emotion
Learning and Explainable AI in Healthcare Using EHR, aware healthcare systems,” Journal of Medical and Biological
pp. 129–148, Springer, New Y ork, NY, USA, 2020. Engineering, vol. 40, pp. 149–157, 2020.
[37] T. Han, N. Stone-Weiss, J. Huang, A. Goel, and A. Kumar, [52] M. Zeraatpisheh, E. Bakhshandeh, M. Hosseini, and
“Machine learning as a tool to design glasses with controlled S. M. Alavi, “Assessing the effects of deforestation and in-
dissolution for healthcare applications,” Acta Biomaterials, tensive agriculture on the soil quality through digital soil
vol. 107, pp. 286–298, 2020. mapping,” Geoderma, vol. 363, Article ID 114139, 2020.
[38] A. Subudhi, M. Dash, and S. Sabut, “Automated segmen- [53] X. Du, P. Wang, L. Fu, H. Liu, Z. Zhang, and C. Yao,
tation and classification of brain stroke using expectation- “Determination of chlorpyrifos in pears by Raman spec-
maximization and random forest classifier,” Biocybernetics troscopy with random forest regression analysis,” Analytical
and Biomedical Engineering, vol. 40, no. 1, pp. 277–289, 2020. Letters, vol. 53, no. 6, pp. 821–833, 2020.
[39] A. Javadi, A. Khamesipour, F. Monajemi, and [54] J. Wang, R. Zuo, and Y. Xiong, “Mapping mineral pro-
M. Ghazisaeedi, “Computational modeling and analysis to spectivity via semi-supervised random forest,” Natural Re-
predict intracellular parasite epitope characteristics using sources Research, vol. 29, no. 1, pp. 189–202, 2020.
random forest technique,” Journal of Public Health, vol. 49, [55] S. Liu, X. Qian, H. Wan, Z. Ye, S. Wu, and X. Ren, “NPC
no. 1, p. 125, 2020. three-level inverter open-circuit fault diagnosis based on
[40] T. Shaikhina, D. Lowe, S. Daga, D. Briggs, R. Higgins, and adaptive electrical period partition and random forest,”
N. Khovanova, “Decision tree and random forest models for Journal of Sensor and Actuator Networks, vol. 2020, Article
outcome prediction in antibody incompatible kidney ID 9206579, 18 pages, 2020.
transplantation,” Biomedical Signal Processing and Control, [56] S. S. Rathore and S. Kumar, “An empirical study of ensemble
vol. 52, pp. 456–462, 2019. techniques for software fault prediction,” Applied Intelli-
[41] K. K. Singh, S. Kumar, P. Dixit, and M. K. Bajpai, “Kalman gence, pp. 1–30, 2020.
filter based short term prediction model for COVID-19 [57] T. Ahmad and H. Chen, “Nonlinear autoregressive and
spread,” Applied Intelligence, pp. 1–13, 2020. random forest approaches to forecasting electricity load for
[42] S.-J. Na, J.-W. Shin, S.-H. Eom, and E.-H. Lee, “A study on utility energy management systems,” Sustainable Cities and
random forest-based estimation model for changing the Society, vol. 45, pp. 460–473, 2019.
automatic walking mode of above knee prosthesis,” The [58] S. Gupta, J. Sarkar, M. Kundu, N. R. Bandyopadhyay, and
Journal of IKEEE, vol. 24, no. 1, pp. 9–18, 2020. S. Ganguly, “Automatic recognition of SEM microstructure
[43] M. Alloghani, T. Baker, D. Al-Jumeily, A. Hussain, and phases of steel using LBP and random decision forest
J. Mustafina, and A. J. Aljaaf, “Prospects of machine and deep operator,” Measurement, vol. 151, Article ID 107224, 2020.
learning in analysis of vital signs for the improvement of [59] L. T. T. Ho, L. Dubus, M. De Felice, and A. Troccoli, “Re-
healthcare services,” Nature-Inspired Computation In Data construction of multidecadal country-aggregated hydro
Mining And Machine Learning, pp. 113–136, Springer, New power generation in Europe based on a random forest
York, NY, USA, 2020. model,” Energies, vol. 13, no. 7, p. 1786, 2020.
[44] Y. Zhu, W. Xu, G. Luo, H. Wang, J. Yang, and W. Lu, [60] Y. Zhou, S. Li, C. Zhou, and H. Luo, “Intelligent approach
“Random Forest enhancement using improved Artificial based on random forest for safety risk prediction of deep
Fish Swarm for the medial knee contact force prediction,” foundation pit in subway stations,” Journal of Computing
Artificial Intelligence in Medicine, vol. 103, p. 101811, 2020. in Civil Engineering, vol. 33, no. 1, Article ID 05018004,
[45] H. Zhang et al., “Deep multi-model cascade method based on 2019.
CNN and random forest for pharmaceutical particle de- [61] X. DengY. Zhan et al., “Predictive geographical authenti-
tection,” IEEE Transactions on Instrumentation and Mea- cation of green tea with protected designation of origin using
surement, vol. 69, no. 9, pp. 7028–7042, 2020. a random forest model,” Food Control, vol. 107, Article ID
[46] H. Lee and E. Jung, “An Analysis of Annual Changes on the 106807, 2020.
Determining Factors for Teacher Attachment with Random [62] S. A. Liu, P. Ngare, and D. Ikpe, “Probabilistic forecasting of
Forest, pp. 463–470, Springer, New York, NY, USA, 2020. crop yields via quantile random forest and Epanechnikov
18 Computational Intelligence and Neuroscience

Kernel function,” Agricultural and Forest Meteorology, [78] S. N. Singh and T. Sarraf, “Sentiment analysis of a product
vol. 280, Article ID 107808, 2020. based on user reviews using random forests algorithm,” Data
[63] H. J. Jeong and M. H. Kim, “Utilizing adjacency of colleagues Science & Engineering, vol. 32, pp. 112–116, 2020.
and type correlations for enhanced link prediction,” Data & [79] H. Zhang, N. Cheng, Y. Zhang, and Z. Li, “Label flipping
Knowledge Engineering, vol. 125, Article ID 101785, 2020. attacks against Naive Bayes on spam filtering systems,”
[64] Z. Khorshidpour, S. Hashemi, and A. Hamzeh, “Evaluation Applied Intelligence, 2021.
of random forest classifier in security domain,” Applied [80] R. P. Kaur, M. Kumar, and M. K. Jindal, “Newspaper text
Intelligence, vol. 47, no. 2, pp. 558–569, 2017. recognition of Gurumukhi script using random forest
[65] J. Tian, L. Liu, F. Zhang, Y. Ai, R. Wang, and C. Fei, “Multi- classifier,” Multimedia Tools and Applications Journal,
domain entropy-random forest method for the fusion di- pp. 1–14, 2019.
agnosis of inter-shaft bearing faults with acoustic emission [81] S. Madichetty and M. Sridevi, “A novel method for identi-
signals,” Entropy, vol. 22, no. 1, p. 57, 2020. fying the damage assessment tweets during disaster,” Futur.
[66] B. Shaw, A. K. Suman, and B. Chakraborty, Wine Quality Gener. Comput. Syst.vol. 116, pp. 440–454, 2020.
Analysis Using Machine Learning, pp. 239–247, Springer, [82] A. Madasu and S. Elango, “Efficient feature selection tech-
New York, NY, USA, 2020. niques for sentiment analysis,” Multimedia Tools and Ap-
[67] K. Madhumathi and T. Suresh, Node Localization in Wireless plications, vol. 79, no. 9-10, pp. 6313–6335, 2020.
Sensor Networks Using Multi-Output Random Forest [83] A.-C. Chang, C. V. Trappey, A. J. C. Trappey, and
Regression, pp. 177–186, Springer, New York, NY, USA, L. W. L. Chen, “Web mining customer perceptions to define
2020. product positions and design preferences,” International
[68] Y. Fang, Y. Xu, C. Huang, L. Liu, and L. Zhang, “Against Journal on Semantic Web and Information Systems, vol. 16,
malicious SSL/TLS encryption: identify malicious traffic no. 2, pp. 42–58, 2020.
based on random forest,” in Proceedings of the Fourth In- [84] R. Kumar and J. Kaur, “Random forest-based sarcastic tweet
ternational Congress on Information And Communication classification using multiple feature collection,” in Multi-
Technology, pp. 99–115, London, UK, 2020. media Big Data Computing For IoT Applications,
[69] T. T. Bhavani, M. K. Rao, and A. M. Reddy, “Network in- pp. 131–160, Springer, New York, NY, USA, 2020.
trusion detection system using random forest and decision [85] A. Onan and M. A. Toçouglu, “Weighted word embeddings
tree machine learning techniques,” in Proceedings of the First and clustering-based identification of question topics in
International Conference On Sustainable Technologies For MOOC discussion forum posts,” Computer Applications in
Engineering Education, 2020.
Computational Intelligence, pp. 637–643, London, UK, 2020.
[86] O. M. Baez-Villanueva and M. Zambrano, “RF-MEP: a novel
[70] P. S. Chaithanya, M. R. G. Raman, S. Nivethitha, K. S. Seshan,
Random Forest method for merging gridded precipitation
and V. S. Sriram, “An efficient intrusion detection approach
products and ground-based measurements,” Remote Sensing
using enhanced random forest and moth-flame optimization
of Environment, vol. 239, Article ID 111606, 2020.
technique,” Computational Intelligence In Pattern Recognition,
[87] A. Beck, “Sentiment analysis on massive open online course
pp. 877–884, Springer, New York, NY, USA, 2020.
evaluations: a text mining and deep learning approach,”
[71] Z. Mingjing, “A novel strategy for quantitative analysis of soil
Computer Applications in Engineering Education, 2020.
pH via laser-induced breakdown spectroscopy coupled with
[88] A. Onan, “Sentiment analysis on product reviews based on
random forest,” Plasma Science Technology, vol. 22, no. 7, weighted word embeddings and deep neural networks,”
p. 74003, 2020. Computer Applications in Engineering Education, Article ID
[72] M.-H. Lee, “Robust random forest based non-fullerene or-
e5909, 2020.
ganic solar cells efficiency prediction,” Organic Electronics, [89] A. Onan and M. A. Tocoglu, “A term weighted neural
vol. 76, Article ID 105465, 2020. language model and stacked bidirectional LSTM based
[73] J. Zhang, G. Ma, Y. Huang, J. sun, and F. Aslani, “Modelling framework for sarcasm identification,” IEEE Access, vol. 9,
uniaxial compressive strength of lightweight self-compacting pp. 7701–7722, 2021.
concrete using random forest regression,” Construction and [90] C. Rodriguez-Pardo, M. A. Patricio, A. Berlanga, and
Building Materials, vol. 210, pp. 713–719, 2019. J. M. Molina, “Machine Learning for Smart Tourism and
[74] W. Nener, C. Wu, H. Zhong, Y. Li, and L. Wang, “Prediction Retail, pp. 311–333, IGI Global, 2020.
of undrained shear strength using extreme gradient boosting [91] W. Song and Y. Zhou, “Road travel time prediction method
and random forest based on Bayesian optimization,” Geo- based on random forest model,” in Smart Trends In Com-
science Frontiers, vol. 12, no. 1, pp. 469–477, 2020. puting And Communications, pp. 155–163, Springer, New
[75] P. Zhang, Z.-Y. Yin, Y.-F. Jin, and T. H. T. Chan, “A novel York, NY, USA, 2020.
hybrid surrogate intelligent model for creep index prediction [92] A. Jamatia, U. Baidya, S. Paul, S. DebBarma, and S. Dey,
based on particle swarm optimization and random forest,” “Rating prediction of tourist destinations based on super-
Engineering Geology, vol. 265, p. 105328, 2020. vised machine learning algorithms,” Computational Intelli-
[76] K. Shah, H. Patel, D. Sanghvi, and M. Shah, “A comparative gence In Data Mining, pp. 115–125, Springer, New York, NY,
analysis of logistic regression, random Forest and KNN USA, 2020.
models for the text classification,” Augmented Human Re- [93] C. F. Baumeister, T. Gerstenberg, T. Plieninger, and
search, vol. 5, no. 1, pp. 1–16, 2020. U. Schraml, “Exploring cultural ecosystem service hotspots:
[77] S. Hakak, M. Alazab, S. Khan, T. R. Gadekallu, linking multiple urban forest features with public partici-
P. K. R. Maddikunta, and W. Z. Khan, “An ensemble ma- pation mapping data,” Urban Forestry & Urban Greening,
chine learning approach through effective feature extraction vol. 48, Article ID 126561, 2020.
to classify fake news,” Future Generation Computer Systems, [94] J. Evans, B. Waterson, and A. Hamilton, “Forecasting road
vol. 117, pp. 47–58, 2021. traffic conditions using a context-based random forest
Computational Intelligence and Neuroscience 19

algorithm,” Transportation Planning and Technology, vol. 42,

no. 6, pp. 554–572, 2019.
[95] L. Zhou, X. Dang, Q. Sun, and S. Wang, “Multi-scenario
simulation of urban land change in Shanghai by random
forest and CA-Markov model,” Sustainable Cities and So-
ciety, vol. 55, Article ID 102045, 2020.
[96] H. Liang, Z. Guo, J. Wu, and Z. Chen, “GDP spatialization in
Ningbo City based on NPP/VIIRS night-time light and
auxiliary data using random forest regression,” Advances in
Space Research, vol. 65, no. 1, pp. 481–493, 2020.
[97] Z. Mei, W. Ding, C. Feng, and L. Shen, “Identifying com-
muters based on random forest of smartcard data,” IET
Intelligent Transport Systems, vol. 14, no. 4, pp. 207–212,
2020.
[98] Q. Li, L. Chen, X. Li et al., “A progressive random forest-
based random walk approach for interactive semi-automated
pulmonary lobes segmentation,” International Journal of
Machine Learning and Cybernetics, vol. 11, pp. 2221–2235,
2020.
[99] S. L. S. Darshan and C. D. Jaidhar, “An empirical study to
estimate the stability of random forest classifier on the hybrid
features recommended by filter based feature selection
technique,” International Journal of Machine Learning and
Cybernetics, vol. 11, no. 2, pp. 339–358, 2020.
[100] S. K. Lakshmanaprabu, K. Shankar, M. Ilayaraja, A. W. Nasir,
V. Vijayakumar, and N. Chilamkurti, “Random forest for big
data classification in the internet of things using optimal
features,” International Journal of Machine Learning and
Cybernetics, vol. 10, no. 10, pp. 2609–2618, 2019.
[101] P. Liu, X. Wang, L. Yin, and B. Liu, “Flat random forest: a
new ensemble learning method towards better training ef-
ficiency and adaptive model size to deep forest,” Interna-
tional Journal of Machine Learning and Cybernetics, vol. 11,
pp. 2501–2513, 2020.
[102] M. Goodwin and A. Yazidi, “Distributed learning automata-
based scheme for classification using novel pursuit scheme,”
Applied Intelligence, vol. 50, no. 7, pp. 2222–2238, 2020.
[103] K. S. Narendra and M. A. L. Thathachar, Learning Automata:
An Introduction, Courier Corporation, Chelmsford, CA,
USA, 2012.
[104] A. Rezvanian, A. M. Saghiri, S. M. Vahidipour,
M. Esnaashari, and M. R. Meybodi, Recent Advances In
Learning Automata, Vol. vol. 754, Springer, New York, NY,
USA, 2018.
[105] R. S. Sutton and A. G. Barto, Reinforcement Learning: An
Introduction, 2017.
[106] B. Pang and L. Lee, “Seeing stars: exploiting class relation-
ships for sentiment categorization with respect to rating
scales,” 2005, http://arxiv.org/abs/0506075.
[107] A. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and
C. Potts, “Learning word vectors for sentiment analysis,” in
Proceedings Of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies,
pp. 142–150, Buffalo, NY, USA, 2011.
[108] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up? senti-
ment classification using machine learning techniques,”
2002, http://arxiv.org/abs/0205070.
[109] D. Dua and C. Graff, UCI Machine Learning Repository,
University of California, School of Information and Com-
puter Science, Irvine, CA, 2019.
[110] C. López-Vázquez and E. Hochsztain, “Extended and
updated tables for the Friedman rank test,” Communications
in Statistics - Theory and Methods, vol. 48, no. 2, pp. 268–281,
2019.