ARXIV.
ORG » COMPUTER SCIENCE » MACHINE LEARNING
A Machine Learning Tool to Determine State
of Mind and Emotion
Rodrigo S. Jamisola Jr.
Abstract—This paper investigates the possibility of creating a
machine learning tool that automatically determines the state
of mind and emotion of an individual through a questionnaire,
arXiv:2108.03444v1 [cs.LG] 7 Aug 2021
without the aid of a human expert. The state of mind and
emotion is defined in this work as pertaining to preference,
feelings, or opinion that is not based on logic or reason. It
is the case when a person gives out an answer to start by
saying, "I feel...". The tool is designed to mimic the expertise
of a psychologist and is built without any formal knowledge
of psychology. The idea is to build the expertise by purely
computational methods through thousands of questions collected
from users. It is aimed towards possibly diagnosing substance
addiction, alcoholism, sexual attraction, HIV status, degree of
commitment, activity inclination, etc. First, the paper presents
the related literature and classifies them according to data
gathering methods. Another classification is created according
to preference, emotion, grouping, and rules to achieve a deeper
interpretation and better understanding of the state of mind and
emotion. Second, the proposed tool is developed using an online
addiction questionnaire with 10 questions and 292 respondents.
In addition, an initial investigation on the dimension of addiction
is presented through the built machine learning model. Machine
learning methods, namely, artificial neural network (ANN) and
support vector machine (SVM), are used to determine a true or
false or degree of state of a respondent.
Index Terms—Diagnostic tool, mind and emotion, compu-
tational psychology, neural network, support vector machine,
addiction questionnaire.
I. Introduction
Machine learning tools are not as widely used in psy-
chology as in health and medicine. This is asserted by the
fact that the interaction between cognition and emotion is
not yet fully understood [1]. Machine learning in medical
applications helped characterize genes and viruses [2]–[6]; [7],
[8], evaluate tumors and cancer cells [9]–[14]; [15], analyze
medical images, [16]–[21]; [22], and assess the health status of
patients [23]–[26]; [15], [27]. Other recent studies in machine
learning include [28]–[35].
A machine learning tool is proposed to mimic the expertise
of a psychologist in determining the state of mind and emotion
of an individual. In this paper, state of mind and emotion
is referred to as something that is not based on conscious
reasoning but is based on one’s feeling or intuition. Thus
normally, when we use it to give our preference or opinion,
we start our statement by saying "I feel ...". The proposed
method presented in this work does not claim any theoretical
contribution to machine learning theory and is purely applied
research. It investigates the possibility of duplicating the exper-
tise of a psychologist through purely numerical computations,
without any formal knowledge of psychology. It is based on
R.S. Jamisola Jr. e-mail: rjamisolajr@gmail.com, jamisolar@biust.ac.bw
1
an online questionnaire, with inputs taken from online users
and are analyzed using machine learning methods. Assessment
questionnaires are extensively used in psychology and are
analyzed by psychologists.
There are several advantages to the proposed approach.
Firstly, the proposed tool can possibly replace the required
expertise in performing an intelligent psychological evaluation.
Secondly, a huge database of questions can be created such
that a fresh set of questions can be provided for retakers.
Thirdly, questions can be designed to be user-friendly in order
to capture a fast response or to avoid a respondent from
intentionally hiding a truthful answer. And lastly, through the
proposed method, it is possible to determine the dimension
of the state of mind and emotion by identifying critical
questions that greatly influence the final output. It is noted
that manual evaluation of this dimension can be very difficult
to determine. This method can possibly lead towards a deeper
understanding of the state of mind and emotion, without the
aid of a psychologist, but through a wealth of questions and
their classifications stored in a repository.
The author recognized the fact that questionnaire-based
diagnosis cannot be very accurate and precise. Issues on
accuracy can occur because respondents can lie, and issues
on precision can arise because even the respondent cannot
be precise about his or her own feelings [36]. (Subsequent
references to “his or her” or “he or she” are omitted for
simplification and are replaced by references to male sex
only to refer to both sexes.) However, the same challenges
are faced by questionnaire-based diagnostic examinations,
whether numerically or professionally analyzed.
There are other advantages of this numerical analysis com-
pared to the human-analyzed questions. Firstly, data errors
in creating the model can be compensated by a statistically
higher number of consistencies in the majority of the gathered
information. Secondly, analysis errors are consistent with the
model and can be easily corrected by reconstructing the model.
Compared to the manually analyzed questions, human error
can contribute to errors in analysis. And thirdly, updating the
model can be fast by removing erroneous data, adding newly
gathered data, and reconstructing the model.
It has long been suggested that machine learning models can
provide better classification accuracy than explicit knowledge
acquisition techniques [37]. Thus in the past two decades,
a considerable number of researches were done in machine
learning and has been applied to a wide range of fields of
study. However, a more recent study by [38] showed that
an analytic instrument from empirical psychometric research
can also prove to be a valid alternative to machine learning
to detect public sentiment. In some cases, machine learning
2 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING
QB
0 review paper [46] pointed out that computational psychiatry
DM
ML 1 tion, improve the selection of treatment or predict the outcome
UI
R selection and predicts drinking patterns using cluster analysis
CA
Fig. 1: A diagram of the discussion presented in this paper.
The center circle represents the machine learning (ML)
method used for classification. Inputs discussed are
questionnaire-based (QB), data mining (DM), user interface
(UI), and camera (CA). The output of classification can be
true (1), false (0), or a number range (R) representing an
extent of influence.
tools are used to solve traditional mathematical computations
[39] which proved to be comparable to traditional results.
Interestingly, the idea of a gaze sensor that has the ability
to detect staring, similar to that of humans was first discussed
in [40].
To analyze the state of mind and emotion, there are two
approaches used in this paper. The first is via a thorough dis-
cussion and analysis of related literature, and the second is via
a machine learning model built on an addiction survey. In the
first part of the paper, previous studies are classified according
to data gathering methods to establish the different modes of
collecting information on choices by respondents that are not
based on reason or logic. This will introduce the reader to
the wide range of mediums that the information on the state
of mind and emotion is collected, whether the respondent has
directly or indirectly provided the information. Then the type
of choices is classified to look into their commonality and
differences in order to gain a deeper interpretation and better
understanding of such choices. The second part is dedicated
to an initial attempt to build a machine learning model of
addiction, which is identified as a platform to investigate the
state of mind and emotion. It also investigates the dimension
of addiction by verifying the independence or interdependence
of the responses to the survey questions.
From the extensive literature gathered, four data gather-
ing methods are identified namely, questionnaire-based, data
mining, user interface, and camera. Fig. 1 shows a diagram
of Machine Learning (ML) discussion presented in this pa-
per. Data-gathering methods are shown as blocks on the
left-hand side: questionnaire-based (QB), data mining (DM),
user interface (UI), and camera (CA). Possible outputs of
machine learning analysis are true (1), false (0) or number
range (R) indicating the extent of influence. More recent
studies in machine learning include a review of probabilistic
machine learning [41], human-in-the-loop [42], a review of
recommender systems [43], and computational nature of social
systems [44].
Lastly, the following recent advances involve machine learn-
ing, questionnaire, or clinical assessment related to addiction.
A 16-scale self-report questionnaire that assesses a range of
addictive behaviors [45] uses traditional statistical analysis and
does not use machine learning methods for classification. A
uses machine learning methods to improve disease classifica-
of treatment. A study identifies risk factors using feature
[47]. It used machine learning to classify from clinical data,
but does not use online questionnaires or identify independent
dimensions of addiction. And, big data has been recognized
unprecedented opportunity to track and analyze behavior [48].
II. Data-Gathering Methods
This study proposes a questionnaire-based data-gathering
method in building a machine learning model. To evaluate
this method, different modes of data gathering in determining
the state of mind and emotion are presented here to show
how they are used. These methods may present advantages as
well as limitations in output classification. From the gathered
literature, four types of data-gathering methods are identified:
questionnaire-based, online data mining, user interface, and
camera. A quick-glance summary of the methods used is
shown in Table I.
In addition, the data gathering methods presented here may
be classified into two types: one with direct user interaction,
and another with indirect interaction. Direct user interaction
includes user interface and camera. In this case, user response
is immediately received while interacting with the machine
learning tool. This is normally done in real-time with at least
one sensor involved. On the other hand, indirect user inter-
action includes questionnaire-based and online data mining,
where the user response is saved and analyzed. Normally this
is not done in real-time and no sensors are involved. This type
of user interaction only requires regular office equipment and
is thus cheaper to implement.
A. Questionnaire-Based
A questionnaire, also known as a survey, is used to gather
information from respondents. This is normally used to as-
sess consumer satisfaction with products and services. One
study used choice-based conjoint analysis that built models
of consumer preferences over products with answers gathered
from questionnaires [49]–[51]. This was a marketing research
technique that was used to determine the required features
of a new product based on feedback from consumers. Two
machine learning tools were used: hierarchical Bayes analysis
and Support Vector Machine (SVM).
Another questionnaire-based study classified students for
an intelligent tutoring system in an adaptive pre-test using a
machine learning tool [52]. Students were profiled based on
performance measurements and gaming preferences through a
questionnaire using Bayesian network and logistic regression
models [53]. In some cases, open-answer questionnaires [54]
were designed to use rule learning and correspondence anal-
ysis to automatically gather useful information. The authors
argued that answers to open-ended questions often contain
valuable information and provide an important basis for busi-
ness decisions. This information included characteristics for
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION
individual analysis targets and relationships among the targets.
A similar approach of information gathering was used in
scoring open-ended responses to video prompts designed to
assess Math teachers [55] using naïve Bayes.
Questionnaires to assess occupational exposure [56] were
used to identify the underlying rules from responses through
regression trees and random forests. In one study by [57],
the gathered data was used to create efficient decision rules
for extracting relevant information from noisy questionnaire
data. It used both simulated breeding (genetic algorithm) and
inductive learning techniques. Simulated breeding was used
to get the effective features from the questionnaire data and
inductive learning was used to acquire simple decision rules
from the data. Through the use of questionnaires before and
after deployment, one study [58] predicted post-traumatic
stress disorder of Danish soldiers after a deployment. It used a
Markov boundary feature selection algorithm and classification
used SVM.
B. Online Data Mining
Online data mining involved an automatic gathering of
information from online content. This is normally performed
by applications that crawl through them and gather information
based on keywords found. The mined data may carry informa-
tion about personal sentiments, opinions, or preferences. This
data was also used to track user’s behavior online and assess
a person’s response or performance.
On sentiment analysis, one study [61] considered sentiment
classification of online reviews as a class of web-mining
techniques that performed an analysis of opinion on travel
destinations. The authors gathered information from travel
blogs, then used three supervised machine learning techniques,
namely, naïve Bayes, SVM, and the character-based N-gram
mode to come up with sentiment classification. Another study
analyzed sentiments on Facebook comments [62], and used
decision trees, naïve Bayes, and SVM to classify them. A
user’s behavior online was tracked through a user’s browsing
history in hypertext [63]. This study involved applying ma-
chine learning algorithms to generate personalized adaptation
of hypertext systems. Conceptual clustering and inductive ma-
chine learning algorithms were used. Predefined user profiles
were replaced with a dynamic user profile-building scheme in
order to provide individual adaptation. A homemade access
log database was used, together with a number of statistical
machine learning models, to compare different classification
or tracking of user navigation patterns for closed world hy-
permedia [64]. Neural network and Markovian models were
used in dealing with temporal data. Another study exploited
the rich user-generated location contents in location-based
social networks [65] to offer tourists the most relevant and
personalized local venue recommendations using the Bayesian
approach.
In searching for online help, users may ask questions that
can be frequently asked questions (FAQ) or not. A study iden-
tified FAQs from non-FAQs [66] by using machine learning-
based parsing and question classification. The authors noted
that the identification of specific question features was the key
3
to obtaining an accurate FAQ classifier. The SVM method and
naïve Bayes were used. Risk-of-bias assessment can be a very
critical factor in systematic reviews. One study tackled this
issue [68] and created three risk-of-bias assessment properties:
sequence generation, allocation concealment, and blinding.
The approach used supervised machine learning and logistic
regression.
Online data mining also considered performance assessment
in classifying credit scores and telemarketing success. In
classifying credit scores, one study [70] used feature selection
algorithms and ensemble classifiers that include principal com-
ponent analysis, genetic algorithm, artificial neural network,
and AdaBoost. Telemarketing success [71] predicted the suc-
cess of telemarketing calls for selling bank long-term deposits
using logistic regression, decision trees, neural networks, and
SVM.
C. User Interface
Previous studies in user interface data gathering can be
classified into two types: one through user input and another is
by detection of brain signals. The first type of user interface
allowed the respondent to input his reaction to a stimulus,
normally through a screen display, by natural language, or
physical cues. The second type is through the detection of
brain signals, normally through electroencephalogram (EEG),
which allowed real-time interaction with the respondent and
online machine learning analysis. It can be used to detect
human emotion, infer user preference, and as a feedback
mechanism to machine learning systems.
Emotion recognition from body movements [72] used cam-
eras and motion tracker sensors to track body movements.
To classify the emotion of the user, the study used Principal
Component Analysis (PCA), naïve Bayes, and Markov model.
Emotion classification by speech [73] was studied where a
speech emotion retrieval system aimed to detect a subset of
data with specific expressive content. The experiment used
a speech sensor to collect data and SVM for emotional
classification. A study used many types of sensors, including
temperature sensors to detect ambient temperature, and to
collect data [74] in a patent application to determine emotion.
Data acquisition devices include a camera, a microphone, an
accelerometer, a gyroscope, a location sensor, and a tempera-
ture sensor to detect ambient temperature. This study outputted
emotion classification using PCA and SVM.
A study was performed to detect human emotion from phys-
iological cues using four machine learning methods: k-nearest
neighbor, regression tree, Bayesian network, and SVM [75].
The respondents interact with computers, and their emotions
were detected by sensors attached to their bodies. Results
showed that SVM gave the best classification accuracy even
though all the methods performed competitively. ID3 (Iterative
Dichotomiser 3) algorithm is used in machine learning and
natural language processing domains. For the study in [76], the
learner’s emotional reaction in a distant learning environment
is inferred using the ID3 algorithm. A study by [36] used
physiological signals to gather data from a single subject over
six weeks. A computer-controlled prompting system called
4 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING

a
TABLE I: Data-Gathering Methods used in Machine Learning Diagnostic Tools
Purpose Technique Output Reference
A. Questionnaire Based
Conjoint analysis Bayes, SVM consumer feedback [49], [50]
[51]; [59]
Student classification Bayesian net., SVM categorized abilities [52], [53]
Analyze open answers rule learning, cor. analysis, naïve Bayes classification rules [54], [55]
Extract relevant data sim. breeding, ind. learning, rand. forests decision rules [56], [57]
Assess traumatic disorder Markov feature selection, SVM stress classification [58]; [60]
B. Online Data Mining
Analyze sentiments naïve Bayes, SVM, N-gram, dec. trees sentiment classification [61], [62]
Track behavior online clustering, neural network, Bayesian dynamic user profile [63], [64]
[65]
Identify FAQ or non-FAQ SVM, naïve Bayes question classification [66]; [67]
Assess review bias supervised ML, logistic regression bias assessment [68]; [69]
Assess performance PCA, neural network, SVM, dec. trees credit score, marketing [70], [71]
C. User Interface
Detect human emotion ID3, k-n.n., SVM, Bayesian net., reg. tree emotion classification [72], [73]
[74], [75]
[36], [76]
Infer user preference decision trees, HMM preference classification [77], [78]
[79], [80]
Feedback to ML systems rule learning, naïve Bayes suggested features [81]; [82]
D. Camera
Detect real-time emotion LDA, AdaBoost, SVM, Bayesian net. emotion classification [83], [84]
[85], [86]
Video facial expression neurofuzzy, Markovian, naïve Bayes emotion classification [87], [88]
3D facial expression LDA 3D facial database [89]; [90]

a
Abbreviated words meaning: net. – network, cor. – correspondence, sim. – simulated, ind. – inductive, rand. – random, dec. – decision, n.n. – nearest
neighbor, and reg. – regression.

“Sentograph” showed a set of personally-significant imagery
to help elicit eight emotional states, namely, no emotion (neu-
tral), anger, hate, grief, platonic love, romantic love, joy, and
reverence. Transforming techniques used sequential floating
forward search, Fisher projection, and a hybrid of the two.
Classifiers used k-nearest-neighbor and maximum a posteriori.
One study [91] considered a prediction of emotional perceptive
competency and implicit affective preferences. It gathered data
through eye-tracking and neurocognitive processes comprising
of six domains: executive function and attention, language,
memory and learning, sensorimotor, visuospatial processing,
and social perception. They were used to predict emotion
through linear regression, PCA with linear regression, and
SVM.
The study in [77] considered inferring interface design
preferences from the user’s eye-movement behavior using an
eye tracker device. Machine learning information processing
is done via decision trees, and this outputs user design prefer-
ences. Folk music preference was studied in [78] using EEG
signals to collect data from the user. SVM classifier with radial
basis function (RBF) kernel was used, and the output is a pre-
dicted user preference. Another study in user preference was
[79] that considered aesthetic preference recognition of 3D
shapes. It gathered user information through EEG signals and
used SVM and k-nearest neighbors to process the information.
A user interface has been devised so different learner prefer-
ences can be acquired through interaction with the system [80].
Based on this information, user interfaces were customized to
accommodate a learner’s preference in an intelligent learning
environment. User preference was diagnosed using decision
tree and hidden Markov model (HMM) approaches.
One study used respondents to communicate feedback to
machine learning systems [81], with the purpose of improving
its accuracy. Users were shown explanations of machine learn-
ing predictions and were asked to provide feedback. These
include suggestions for re-weighting of features, proposals for
new features, feature combinations, relational features, and
changes to the learning algorithm. Two learning algorithms
were used: the Ripper rule-learning algorithm and the naïve
Bayes algorithm. The study showed the potential of rich
human-computer collaboration via on-the-spot interactions, to
share intelligence between user and machine.
D. Camera
The last method discussed in the paper for data gathering
is through the use of a camera. This method can perform a
real-time observation of facial expression, or can be non-real-
time through a video recording, which the machine learning
method then analyzed to output a judgment. One disadvantage
of relying on face or voice to judge a person’s emotion is that
we may see a person smiling or hear that her voice sounded
cheerful, but this does not mean that she was happy [36].
But because human emotion is greatly displayed by facial
expression, its detection by a camera is extensively studied.
To predict negative emotion, one study [92] made use of
the mobile phone camera and processed the information using
a naïve Bayes classifier, decision tree, and SVM. A study that
used a camera to detect facial expression [83], [84] utilized
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION 5

AdaBoost for feature selection prior to classification by SVM TABLE II: Classifications in Determining
or linear discriminant analysis (LDA). Facial expressions in State of Mind and Emotion
the video were analyzed in a study in [85]. It developed an Purpose References
authentic facial expression database where the subjects showed A. Classification: Preference
natural facial expressions based on their emotional state. Then - Consumer product [49], [50], [93]
it evaluated machine learning algorithms for emotion detec- - Travel destination interest [61], [65], [94]
- Ranking aesthetic preferences [77], [78], [95]
tion including Bayesian networks, SVMs, and decision trees. - Analyze online sentiments [80], [81], [96]
Local binary pattern (LBP) was used for facial expression - Tracking of navigation patterns [63], [64], [97]
recognition. The authors used boosted-LBP to extract the most B. Classification: Emotion
discriminant LBP features, and the results were classified - Detect emotion from physiological cues [36], [74], [75]
- Emotion detection from speech [98]–[100]
via SVM. It was claimed that the method worked in low - Online facial expression from camera [83], [84], [101]
resolutions of face images and compressed low-resolution - Off-line facial expression from video [85], [88], [102]
video sequences captured in real-world environments [86]. C. Classification: Grouping
Extraction of appropriate facial features and identification - Student abilities prediction from response [52], [103], [104]
of the user’s emotional state through the use of a neurofuzzy - Identifying off-task behavior [58], [105], [106]
- Model formation to predict future actions [66], [107]
system was studied [87], which can be robust to variations - Gaming-detection model for tutoring behavior [53], [108]
among different people. Facial animation parameters are ex- D. Classification: Rules
tracted from ISO MPEG-4 video standard. The neurofuzzy - Open answers to questionnaires [54], [55]
analysis was performed based on the rules from facial ani- - Efficient decision rules from noisy data [56], [57], [109]
mation parameters variations both at the discrete emotional - Learning casual relationships, word meanings [110]–[112]
- Production rules from independent assessment [68], [113]
space and 2D continuous activation–evaluation. The multi-
level architecture of a hidden Markov model layer was shown
in [88] for automatically segmenting and recognizing human
facial expressions from video sequences. Classification of ex- A. Correlation Among Classifications
pressions from video used naïve Bayes classifiers and learning
the dependencies among different facial motion features used In this work, preference is referred to as an individual’s intu-
Gaussian tree-augmented naïve Bayes classifiers. itive choice given two or more options. It does not include any
A 3D facial expression recognition was shown in [89] that emotion. For visualization purposes, preference can be thought
has developed 3D facial expression database. It has created of as a “horizontal” expression of one’s feelings, where the
prototypical 3D facial expression shapes and 2D facial textures emotional level remains “flat.” On the other hand, emotion
of 2,500 models from 100 subjects. LDA classifier was used is not a conscious choice but an individual’s reaction to an
to classify the prototypic facial expressions of sixty subjects. outside stimulus that affects the person’s disposition. Emotion
From all the four types of data-gathering methods, one may is not based on intuition because intuition involves a mental
say that the sensor-based inputs with direct user interaction process without conscious reasoning. It can be thought of as a
can be considered to be more accurate compared to the “vertical” expression of feelings with varying intensity. Thus
indirect, non-sensor-based method where the user inputs may the usual reference to “up-and-downs” of emotion. Preference
be subjective. However, sensor-based inputs may be limited to and emotion are direct results from individual responses and
what a sensor can detect. For example, a camera may detect are normally referred to as feelings. In other words, preference
a smiling face, but it does not mean the person is happy. Or is a “non-emotional” feeling and is intuitive, while emotion is
if the sensor is not accurate enough to detect the brain signal, an “emotional” feeling and is not intuitive. Therefore feelings
it can give out other output instead of what the user intended. involve a mental process (intuition) and a non-mental process
Thus if the user is objective in his inputs, the questionnaire- (emotion).
based or data-mining method may be more accurate than the On the other hand, grouping and rules classification are
sensor-based. Until the time that more sophisticated sensors not direct results from individual responses. Rather, individual
are developed to detect accurately what the person really wants responses are further analyzed to output a final judgment.
to convey, at the current technological state, the questionnaire- In grouping, classification rules are applied to the individual
based method may significantly cover the user’s actual state responses to classify them according to a set grouping. In
of mind and emotion. rules, the individual responses are used to create new rules
or modify existing ones, which may later be used to arrive
III. Types of Classifications at a final judgment. In terms of the interaction with the
From the previous section, we were able to analyze the respondents, classifications by rules and grouping normally
method of data gathering in the state of mind and emotion, may entail an indirect interaction, while the emotion and
including the questionnaire-based method that is proposed in preference classifications normally require direct interaction
this work. Using the same literature discussed in the previous and are usually performed in real-time. And lastly, in terms
section, together with a few more additional pieces of liter- of decision outcomes, emotion and preference classification
ature, we propose four ways in classifying the state of mind are normally decided by the user. In grouping and rules
and emotion once data has been gathered. The proposed four classification, the decision outcomes are normally decided by
classifications are preference, emotion, grouping, and rules. an observer. Table II shows the summary of classifications.
6 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING
B. Classification by Preference
Preference is an option chosen by an individual based on
how he feels, but with no emotion attached to the judgment.
It is mostly used to identify the liking of a user to a particular
person, place, product, or service. Traditionally, this method
of gathering preference information from users was used by
many companies [49]–[51], [93] to assess their current market
share or to estimate the degree of acceptance of a new product
introduced to the market.
In recent years, user preference posted online is becoming
a new and powerful approach in gathering and analyzing such
information. One approach was by tracking navigation patterns
online and present the most likely information that will be
of interest to the user based on navigation preference [63],
[64]. It can be used to present the most likely advertisements,
interactive interfaces, or locations of places that will be of
interest to the user. Another approach was tracking user
preference on travel destinations [61], [65] or preference on
aesthetics [77]–[79], [95] or of online sentiments [62], [80],
[81].
One can say that such information is monetarily driven by
companies providing products and services. But this can also
be very helpful to users who might want to find immediate
solutions to urgent needs. Thus nowadays, matching demand to
its solution can be quite easily performed by analyzing online
user preference. The other advantage of posting preference
online is that the online document can become the source of
information for other users. For example, other users can put
additional online reviews for a particular travel destination,
making the expanded information more exhaustive and useful
for potential visitors. This is also true for political sentiments
that gather huge support given a short period of time. This has
been a vehicle of many social actions within the past decade.
Thus online preference can become a powerful tool for users
of the same liking. This enables them to bargain for better
service or initiate a desired social change.
C. Classification by Emotion
Of the four types, emotion classification is quite extensively
studied. It is normally detected through camera or EEG
signals, and it enjoys significant interest among researchers.
Emotion is an expression of the feelings of an individual with
varying intensity according to the degree of feelings conveyed.
Emotion can be transmitted and be strongly shared among
individuals, as in a mob. It has the ability to overpower
all other senses of the individual to assume singularity of
purpose. The tone of emotion can be set given an appropriate
environment, as in a relaxing environment with soft music and
dim lights. Or it can be instantaneously derived by giving the
right stimulus as in the case of anger by striking a sensitive
chord or happiness by watching a cute baby. Normally, there
are three types of emotional stimulus, namely, visual, by
hearing, and by touching. Thus a person can be stimulated
visually as in a movie that is horror, comedy, or sexually
explicit; or stimulated by sound as in a vile language or
shouting or relaxing music; or stimulated by touching as in
shaking hands, hugging or kissing, or strong dislike by hard
grip or punching.
As emotion can be stimulated visually, or by sound, or by
touching, it can also be manifested in the same manner. Thus
from previous studies, visual manifestation of emotion through
facial expression was identified in real-time using camera [83],
[84], [87], [89] or off-line using recorded video [85], [86],
[88]. Emotion was also identified by sound through speech
[73], [98], [99]. The last method may not be obvious from a
human observer when emotion was detected through the use
of sensors attached to the body, and was identified from bodily
cues [36], [72], [74], [75].
D. Classification by Grouping
In grouping classification, the response of individuals are
inputs to the machine learning algorithm which outputs the
group classification. This is different from classifications of
preference and of emotion where the responses are direct
outputs of the classification.
Teachers used classification by grouping to assess students
on the appropriate level of training to be administered [52],
[103], [114]. This gave them an idea of the optimal strategy
to be adopted for each group of students, especially when
a considerable disparity was observed from the grouping
assessment. On the other hand, psychologists used group
classification to assess mental conditions or capabilities [55],
[58], [105] in order to carry out the further intervention, or
to perform an appropriate level of service. Once the group
classification was determined, one will only need to match
a predefined intervention that was appropriate for the corre-
sponding group.
From the groupings, a model of classification can be formed.
This may be a new grouping with new characteristics or an
existing grouping with modified characteristics [66], [107].
This is different from the rules classification because in this
case, the output is groupings and not rules. As the groupings
are formed, the model of the groupings may change. Then
the characteristics are determined based on the new grouping
models, such that the behavior of a group can be predicted.
The last method of grouping classification is very much related
to the students (or training) classification but the method of
determining the grouping was through gaming [53], [108].
This may result in a more appropriate grouping for younger
students because normally they are more alert during a game
interaction, which may help in getting a more accurate re-
sponse from them.
E. Classification by Rules
Lastly, the classification of rules establishes relationships
among different user responses in order to influence a decision-
making process. It does not necessarily output a final classi-
fication but modifies policies or methods that influence the
desired output.
One such classification was through open answers to ques-
tionnaires [54], [55] where classification and association rules
were defined to characterize targets and establish relationships
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION 7

among them. Although the answers were open, from the key- where yi ∈ {1, 0, [10, 100]}. The function f is numerically
words and phrase level, their models can be created, and thus derived from an artificial neural network or support vector
the rules that define their relationships can be formed. Rule machine.
classification was also used in determining underlying rules to
make expert decisions [56], [57] even with noisy data. Rules
A. Artificial Neural Network
were classified in determining the casual relationships and
word meanings [110], [111], in order to understand the idea of Artificial neural network (ANN) has been extensively used
what the person is trying to convey. This can be used in spoken in many different machine learning applications. Two widely
language or from documents to develop models for inductive used types are feedforward multilayer perceptron and radial
learning and reasoning, or from construct psychology. basis function.
Classification of rules for tutoring systems and risk-of- For multilayer perceptron, given input layer i and output
bias assessments were studied in [68], [113]. In the tutoring layer j, a weight between layers i and j is denoted as wi j , such
system, the purpose was to automate rule generation in the that for n nodes in layer i, w j ∈ Rn = [w1 j , . . . , wi j , . . . , wn j ]. An
system development, such that production rules were gen- output of a single node in layer j for a given input x ∈ Rn can
erated from marked examples and background knowledge. be expressed as
In the assessment study, model rules were generated for the n
X
properties of sequence generation, allocation concealment, and yj = wi j xi . (2)
blinding. The models predicted sentences that contain relevant i=1
information, as well as the risk of bias for each research article.
For an input layer i, hidden layers j and k, and a single node
This work has proposed four classifications of the state of
output (output layer l), we can recursively apply (2) three
mind and emotion using the different data-gathering methods
times, to get the input-output relations to be
shown in the previous section. These classifications enable us
to see the different aspects of the state of mind and emotion, Xq

Xp
 n
X


such that its range of forms was discussed. To verify these yl = wkl f   w jk f  wi j xi 
 (3)
different aspects we need to select an experimental platform k=1 j=1 i=1
that enables us to gather data from its range of forms in order
such that wl ∈ Rq , wk ∈ R p , w j ∈ Rn , and y = f (·) is called the
to gain a deeper understanding of its nature. In this work,
activation function.
we choose addiction as the experimental platform. What is
unique to addiction is that it covers all the four classifications For radial basis function with one single output, given input
discussed in this work. x and number of samples m, the following equation can be
applied
Preference, which is a choice by feelings that involves a
mental process or intuition, is performed when people have m
X
manageable addiction. It is normally done when decisions are y= w j φ(k x − x( j) k) (4)
not driven by urges or emotion. On the other hand, choices j=1
made by emotion are normally done by people who have a
where φ(·) is a set of radial basis functions, x( j) is a center of
higher degree of addiction. The choices do not anymore in-
the radial basis function, and w j is an unknown coefficient.
volve a mental process. In terms of classification by grouping,
the choices made by the respondents are analyzed by psy-
chologists that output the groupings. And lastly, classification B. Support Vector Machine Model
by rules that involves the process of looking for new rules Support vector machines (SVM) are derived from statistical
or modifying existing ones applies to addiction because when learning theory [115]. It has two major advantages over other
gathering data via questionnaires, the relationships among the machine learning tools: (1) it does not have a local minimum
questions are verified and modified accordingly. This can lead during learning, and (2) its generalization error does not
towards understanding the dimensions of addiction and can depend on the dimension of the space. Given m samples (xi , yi )
affect its final output classification. Thus addiction indeed where i = 1, . . ., m, for an i-th sample input xi ∈ Rn , a scalar
covers all four classifications discussed in this section. offset b ∈ R and a weighting vector w ∈ Rn , a function f is
given as
IV. Machine Learning Classifiers
f (xi ) = w · xi + b. (5)
Two machine learning classifiers are considered: artificial
neural network and support vector machine. The number of A loss function L that is insensitive to tolerable error ǫ can
questions will be equal to the dimension of the input space, n. be expressed as
For an i-th sample, the corresponding answers can be true (1), m
CX
false (0), or a degree of state ([10, 100]). Thus for an input L = kwk2 + max{0, |yi − f (xi )| − ǫ} (6)
xi ∈ Rn , a function f : Rn → R is defined as m i=1

where C ∈ R is a regularization constant which can be ex-

yi = f (xi ) (1) pressed as an optimization problem in the form
8 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING
1 CX
m questions were asked, with possible answers ‘yes’, ‘no’, and
min kwk2 + (ξi + ξi∗ )
2 m i=1
subject to: (w · xi + b) − yi ≤ ǫ + ξi (7)
yi − (w · xi + b) ≤ ǫ + ξi∗
ξi , ξi∗ ≥ 0 for i = 1, . . . , m.
To test the proposed method, an online addiction ques-
tionnaire with 10 questions was created and answers from
292 respondents were analyzed. Dependence/independence of
questions were verified by removing questions one by one and
noting the resulting accuracy of classification, which can be
further developed to determine the dimensions of addiction.
V. Addiction as an Experimental Platform
To test the proposed machine learning tool, addiction is
used as an experimental platform because it encompasses the
entire range of forms of state of mind and emotion, especially
based on its four classifications stated in the previous section.
Depending on the extent of addiction, the person’s intuitive
response can be consciously or unconsciously made. When
one is not addicted to a stimulant, his choices are consciously
made and he is in total control of his reaction. For the state of
emotion, normally, the emotional reaction is not consciously
controlled but results from an urge or a bodily reaction that
automatically occurs given the right stimulant. That is why
some people easily cry at sad movies or laugh at certain types
of humor. But when one is addicted to a stimulant, the person’s
reaction is based on an urge or an uncontrolled bodily reaction,
similar to the emotional reaction. The person’s choice, in this
case, is based on unconscious preference, and his reaction is
based on bodily urges. Thus, the study on addiction offers a
platform that considers conscious and unconscious decisions,
as well as controlled and uncontrolled reactions. Furthermore,
the intervention by a psychologist to come up with groupings
based on inputs from respondents shows classification by
grouping. And finally, the classification by rules points to
the attempt in identifying the dimensions of addiction by
characterizing the interdependence of inputs from respondents.
VI. Experimental Results
A questionnaire [116] was designed to gather information
regarding respondents’ degree of addiction to an activity. The
questions were all composed by the author who has no formal
training in psychology, thus may be considered as random
questions. This is done in order to mimic the method of
gathering random questions from users that will be included
in the database. These questions do not claim completeness in
addressing all the dimensions of addiction but are presented
in order to show how any given randomly gathered set of
questions are processed and analyzed. The analysis is in terms
of their interdependence, which may possibly lead towards
clustering questions in the database and furthermore, may
possibly lead towards identifying the dimensions of addiction.
At the end of the questionnaire the respondent will rate self
as ‘addicted’, ‘not addicted’, ‘manageable’, and ‘don’t know’.
A total of 292 respondents participated in the survey. Ten
‘not really’, as well as a range of numbers to rate the frequency
of occurrences or number of persons involved. Most of the
respondents are students and staff from the University of the
author where the average age ranges from 20-30 years old.
Information about the sex of respondents was not gathered
because the addiction study in this work is intended to be
independent of this information, including other biases like
culture, educational attainment, sexual orientation, race, etc.
The author envisioned millions of responses from all over
the world that are normalized to any biases due to the
randomness of the respondents. The actual questionnaire and
the percentage of responses are shown in Appendix IX.
A. Experimental Setup
Two machine learning experiments were performed using
Matlab R2017a neural network, and statistics and machine
learning toolboxes. The neural network toolbox used ‘newff’
function to create a feed-forward backpropagation network
with 85 hidden layers, four output layers with four outputs
that represent four classifications. Data division was random
such that from 292 samples, 204 are used for training, 44
for validation, and 44 for testing. The training algorithm used
the Levenberg-Marquardt backpropagation technique, while
the performance measure was by mean squared error. After
the network has been trained, validated, and tested through
‘train’ command, the network was again tested using ‘sim’ the
command that used all the 292 samples, then the output was
compared against known target values. The average accuracy
was at 77%.
The SVM experiment used ‘templateSVM’ function to
create an SVM template that invoked the Gaussian kernel
function. Then ‘fitcecoc’ was called to train the classifier using
the SVM template that was created. The purpose was to group
the responses into four classifications. The training function
used binary learner and one-versus-one coding design. After
the SVM training, 10-fold cross-validation by ‘crossval’ com-
mand was used. Then the command ‘resubPredict’ was used
to predict the classification from all learners. Its results were
compared using the target values from the 292 samples. SVM
accuracy was measured at 85%.
B. Analysis of Answers by Respondents
In this subsection, we are going to analyze the answers of
the respondents based on their own self-assessment of whether
or not they are addicted to activity. Of the 292 respondents,
20% considered themselves addicted to the activity, 23%
not addicted, 45% manageable, and 12% did not know their
status. The case of “manageable” could mean that the person
enjoys the activity but is in total control of his decisions and
reactions to it, and is thus not addicted. On the other hand,
it can also mean that the person is partially addicted, and has
some control over his decisions and reactions to the activity.
Note that an addicted person, as defined above, has totally
uncontrolled decisions and reactions.
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION
On the frequency of performing the activity, 50% answered
“everyday.” However, less than half of this number admitted
addiction to the activity. This result showed that the everyday
performance of an activity that one likes to do does not
necessarily mean addiction to that activity. This also means
that it is possible for the person to enjoy the addictive activity
every day, but is still in control over it. Regarding the urge to
do the activity, 59% admitted to feeling the urge but only one-
third of them considered themselves addicted. This percentage
is higher compared to the percentage of everyday activity,
which means that those who felt the urge, did not necessarily
perform the activity every day. Furthermore, feeling the bodily
urge to do an addictive activity does not easily overcome
conscious actions and decisions.
On non-performance of the activity on a regular basis and
affecting the mood of the respondent, 43% answered “yes”.
But only half of this percentage admitted addiction. This is
interesting because it means that even without being addicted
and in control over the activity, the person can still be affected
in his regular daily work through his moods. It can also
mean that the effort to control the urge to do the activity can
somehow affect the everyday mood of the person. Solitary
performance of an addictive activity, with a response of 50%,
does not necessarily equate to addiction. A higher percentage
felt the urge to do the activity, but this does not necessarily
mean that they are going to do the activity alone.
Having many other major activities besides the addictive
activity can be a possible source of getting one’s focus away
from the addictive activity. However, 66% answered the least
number of other activities at “three more” and 19% answered
“many.” Thus, in this case, the addicted person can have many
other activities besides work and study. This can also mean that
even the not-addicted person has a limited number of activities
besides work and study.
Talking to close friends and family every day as a support
group can be vital in coping with addiction. Around 45%
talked to “one or two” and 36% talked to “three to five.”
Isolation, in this case, does not seem to have a close connection
to addiction, as in the case of performing the activity alone.
This could mean that the person can be having many friends
and seemed to have a normal life, but is addicted. On asking
for professional help to stop the activity, a huge percentage
of 75% answered “no.” This could mean a lack of access to
professional help or the hesitation to admit the need for help.
Distraction from daily work or studies caused by the addic-
tive activity has 26% who answered “yes.” This is close to the
percentage who admitted addiction to the activity. One may
say that in this case, a distraction from daily routine caused by
the addictive activity is a clear indication of addiction. (This
will be supported by the result in the next section showing this
as a critical question.) Talking about the activity to somebody
else as a possible source of support has 36% answered “no, I
keep it to myself.” This scenario of isolating oneself is related
to the question about the solitary performance of the addictive
activity, and to the number of close friends and family that
one talks to every day. It is noted that the performance of the
activity alone has a higher percentage, which means that of
all the persons who may be performing the activity alone, a
9
large percentage of it kept it as a secret.
A number of things can be noted in order to improve the
machine learning results. One has to design the questions
that tackle independent aspects of the psychological state.
This will enhance a clearer separation in the classification.
Another possible approach is to create subtle support questions
to verify consistency in the answers of the respondent, most
especially to critical questions. Lastly, indirect questions can
be designed so as to avoid the respondent explicitly hide the
truthful answer.
C. Investigating the Dimensions of Addiction
This subsection will analyze the dependence/independence
of one question against the rest of the questions based on the
output classification accuracy. Using the trained model, each
question (response) was removed from the input data, and the
accuracy of the output was observed. If the accuracy of the
output drastically reduced in the absence of a given question,
this means the question was critical and was independent of the
rest of the questions. On the other hand, if a given question was
removed and the resulting output did not change drastically,
that means it was dependent on at least one other question.
In other words, that question did not matter. In this study,
a drastic decrease means a 25% reduction from the overall
accuracy.
There were 10 questions that each respondent had to answer,
and question ten (Q10 ) was a self-assessment based on the four
classifications. Using the model that was created, responses to
question one (Q1 ) up to question nine (Q9 ) were removed one
by one from the input data, and the accuracy of the output was
compared against the target values, that is, the responses to
Q10 . The SVM classification model was used in this analysis.
From the overall accuracy of 85%, an accuracy reduction
of 25% is an output accuracy of around 64%. Table III shows
the resulting percentage accuracy of classification when at
most two questions were removed. The diagonal elements
in the table (in boldface) represent the percentage accuracy
when question Qi was removed. (In the table, i = 1, . . ., 9.)
The encircled values show an accuracy reduction of 25% or
more. The percentage accuracy shown in row Qi is the case
when question Qi was removed first and questions Q1 to Q9
were removed second, one at a time, except Qi . Thus using the
convention (row, column) to define the elements in the table,
the percentage accuracy in (Q1 , Q1 ) is the case when only
question one was removed, and (Q1 , Q2 ) is the case when Q1
was removed first and Q2 was removed second.
From Table III, it can be observed that removal of Q8
drastically lowers the output classification accuracy, which
resulted in (Q8 , Q8 ) = 45%. This drastic decrease in accuracy is
generally consistent all throughout the elements of Q8 row and
column, except for (Q5 , Q8 ) = 75% and (Q8 , Q5 ) = 76%. This
could mean that for most elements in Q8 row and column, the
removal of Q8 and one other question, is greatly influenced by
the absence of Q8 alone. The other question did not greatly
affect the accuracy results, except Q5 . We note question eight
below.
Q8 : Do you think you get distracted in your
10 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING

TABLE III: Percentage Accuracy with at Most Two When Q8 was removed, the new accuracy was drastically
Questions Removed reduced to 45%. But when Q5 was removed next, the new
Qi Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 accuracy drastically increased to 76%. This means that Q5
was dependent on Q8 alone such that when Q8 was removed,
Q1 81 76 80 74 78 66 79 49 70
it became an independent question and did affect the accuracy
Q2 77 82 74 60 80 67 78 58 78 drastically. Let’s investigate now the reverse order of removal.
Q3 80 74 82 76 77 74 75 59 73 When Q5 was removed, the accuracy did not change much and
was at its highest value among the rest of the single questions
Q4 70 60 76 79 78 74 80 61 72
removed. But when Q8 was removed, the accuracy did not
Q5 80 80 78 79 83 76 81 73 74 drastically change. This means that the removal of Q5 affected
Q6 66 68 76 75 75 78 74 57 66 Q8 such that it was not able to drastically change the accuracy
Q7 79 78 75 79 81 68 81 55 81
as it did when the other questions were removed. Thus, Q8
was dependent on Q5 . But initially, we identified Q8 to be
Q8 62 71 59 60 76 58 58 45 65 a critical question because it drastically changed the accuracy
Q9 74 77 73 73 74 70 78 60 80 when removed alone. The explanation is that the characteristic
of Q5 was very similar to the overall accuracy such that when
Q8 was removed alone, this dependence was not obvious.
daily work or studies by thinking about this Another observation is the relationship between Q8 and Q2 .
activity? The case of (Q2 , Q8 ) = 59% but (Q8 , Q2 ) = 71%, that is, the
order of removal has an effect on the resulting accuracy. In
This could mean that distraction from the daily activity is the first case, the accuracy did not drastically change when
generally independent of the rest of the questions in the Q2 was removed. The drastic change of 59% happened only
addiction survey questionnaire. Thus, it can be considered when Q8 was removed after Q2 . This is the same case as
a critical question and can be counted as an independent when Q8 was paired with the rest of the questions, except
dimension of addiction. Q5 . Which means that the removal of Q2 did not affect the
Removal of Q2 and Q4 resulted in a more drastic decrease removal of Q8 , and therefore Q8 is independent of Q2 . In the
in accuracy compared to the removal of Q2 or Q4 alone. This second case, removal of Q8 resulted in a drastic decrease of
could mean that both Q2 and Q4 belong to one class of critical accuracy to 45%, but when Q2 was removed after Q8 , the
questions, which are independent of the rest of the questions. accuracy drastically increased to 71%. Thus Q2 was affected
by the removal of Q8 and is therefore dependent on Q8 . But
Q8 is not dependent on Q2 , thus the dependence is only in
Q2 : Do you feel an urge to do it? one direction and not both. This explains why the order of
Q4 : Do you do this activity alone or with some removal has an effect on the resulting accuracy.
company? The approach presented above can be used to identify
critical questions that drastically change the accuracy output
The relationship can be that the feeling of a strong urge to when removed. Critical questions can help identify the number
do an addictive activity is somehow related to doing such of independent variables in the state of mind and emotion
activity on one’s own accord, that is, being alone. And when and can help in determining its dimensionality. Identifying
the urge is lesser, it is somehow related to the performance critical questions can also help in minimizing the questions
of the activity with more company. Thus, we can say that asked in the questionnaire, in order to save time for the
another independent dimension of addiction includes the urge respondents. This possibility of quantifying the dimensionality
of performance or the number of persons involved during the of a person’s state of mind and emotion by an individual with
performance. no sufficient background in psychology upholds the advantage
A peculiar observation is Q5 and Q8 and we note below. of a machine learning tool that can help replace the “expertise”
Question Q5 is stated in the following. required to perform an intelligent evaluation. It is noted that
the identification of this dimensionality can be very difficult
Q5 : Besides work or studies, how many other for an experienced psychologist to discover.
main activities you have in a day besides this Another future direction of the proposed method is the
activity? possibility of developing questions with hidden information
such that the respondent cannot intentionally cheat on his
We note that the removal of both Q5 and Q8 resulted in a answers. In addition, questions or choices of answers can be
higher accuracy compared to the removal of Q8 alone, such designed to capture a faster response from the respondents,
that (Q8 , Q5 ) = 76% and (Q5 , Q8 ) = 73%. That is, the removal such that the questionnaire can be more user-friendly. This
of both Q5 and Q8 resulted in a 10% accuracy reduction way, user-friendliness from the perspective of the respon-
against the overall accuracy, but removal of Q8 alone results in dent can be accommodated without compromising technicality
a 25% accuracy reduction. We note further that the removal of from the psychologist’s perspective. Furthermore, this can
Q5 alone had almost zero percent accuracy reduction, and is in lead to drastically increasing the number of questions in the
fact the highest accuracy that is closest to the overall accuracy. database, such that the addicted person can test himself again
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION
without answering exactly the same set of questions. This can
make the self-assessment more reliable.
VII. Survey Questions vs. SADD Questionnaire
A questionnaire designed by the author to assess the degree
of addiction to an activity by a respondent, called “A Survey
on Addiction,” is shown in Appendix IX. This set of questions
will be compared against a standard psychological test called
“Short Alcohol Dependence Data (SADD)” Questionnaire
[117] that is used to assess oneself to alcohol addiction. And
secondly, we will analyze the answers of the 292 respondents
that participated in the addiction survey.
One major difference in the addiction survey questions in
this paper compared to the SADD questions are that the
questions in this paper assessed addiction before the tangible
effects are experienced. They did not tackle cases about
physical effects of addiction like “shaky hands”, “vomiting”,
“imaginary” things, etc. but these are included in the SADD
questions. In the following discussion, we compare the first
few questions from SADD against the questions of the survey
shown in Appendix IX.
Question one of SADD addresses the issue of getting the
thought of drinking out of the mind, and this is similar
to question eight in the survey which asked regarding the
thought on the addictive activity being distractive to daily
work or studies. Obviously, when something distracts your
daily routine, it means the thought about it is always in the
mind.
Question two of SADD talked about misplaced priorities
due to alcohol addiction. This is related to question five in
this work that talked about major activities including addictive
activity. But the SADD question is transparent in asking about
misplaced priorities. Being transparent in the question can be
an advantage to get a clear answer regarding it. Or this can
be a disadvantage as well when the respondent will try to
suppress from giving an accurate answer. Thus an indirect
question might be able to address this issue.
Question three of SADD where the activities of the respon-
dent are revolving around alcohol drinking is again related
to question five that asked about major activities of the
respondent including the addictive activity. That is, if the
addictive activity constitutes a major activity of the respondent
then the addictive activity greatly influences all the other
activities. Question four of SADD considers the frequency of
drinking alcohol and is related to question one in this work that
explicitly asks about the frequency of performing the addictive
activity. In this question, the two approaches are very closely
related.
Question five of SADD asked about the desire to satisfy the
need for alcohol without considering the quality of the drink
and is related to question two of this work that asked about the
urge to perform the addictive activity. In this way, the urge to
do the addictive activity created the possibility to disregard any
discomfort that may be experienced in performing it. Lastly,
there were questions in SADD asked about the awareness
of possible consequences of drinking alcohol. This is related
to question two of this work that considered the urge to
11
do the addictive activity without considerations of possible
consequences. It can also be related to question eight that
asks about seeking professional help because of the possible
consequences of the addictive activity.
VIII. Conclusion and Future Direction
This paper has shown the possibility of determining the state
of mind and emotion of an individual through a questionnaire-
based machine learning tool, using an artificial neural network
and support vector machine. Previous classifications and data-
gathering methods were presented to determine preference,
opinion, emotion, or capability. The proposed method is im-
plemented in analyzing addiction through a survey on addic-
tion with ten questions. Results analysis showed a proposed
method to identify critical questions that can lead to the
identification of the dimension of addiction. Analysis of the
survey questions against a standard questionnaire on alcohol
addiction is presented. This tool can be used to do the same
method of computation for all applications but will vary only
on the types of questions asked depending on the individual
information to be extracted. The proposed machine learning
diagnostic tool may be able to output judgment, based on
the thousands of inputs collected from users. The future
direction of this research is for a psychologist to assess,
compare and validate the proposed method and its results. In
addition, a deeper investigation of the dimensions of addiction
via a machine learning model will be performed. Lastly, as
the online database of questions and answers increased, it is
recommended to use unsupervised machine learning to build
the state of mind and emotion model through the correlation
of the responses from the respondents.
Acknowledgment
The author would like to acknowledge the contribution
of Mario Saiano, Social Health Educator, Local Health Unit
Genovese, Italy for his inputs in the preparation of this
manuscript.
IX. The Questionnaire on Addiction Survey
A survey is designed to assess the addiction of a respondent
to an activity. The survey was posted online using Google
forms [116]. This section shows the instruction, questions, and
responses from 292 respondents.
Instruction: This survey consists of 10 questions. Please be
honest in answering. Think of one type of activity that you
like, answer the following questions, and judge for yourself at
the end of the survey if you consider yourself addicted or not
to this activity.
References
[1] J. Taylor, K. Scherer, and R. Cowie, “Emotion and brain: Under-
standing emotions and modelling their recognition,” Neural Networks,
vol. 18, no. 4, pp. 313 – 316, 2005, Special Issue: Emotion and Brain.
[2] M. A. Shipp, K. N. Ross, P. Tamayo, A. P. Weng, J. L. Kutok, R. C.
Aguiar, M. Gaasenbeek, M. Angelo, M. Reich, G. S. Pinkus, et al.,
“Diffuse large b-cell lymphoma outcome prediction by gene-expression
profiling and supervised machine learning,” Nature Medicine, vol. 8,
no. 1, pp. 68–74, 2002.
12 ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING

TABLE IV: A Survey on Addiction a neural network classifier,” Information Technology in Biomedicine,
IEEE Transactions on, vol. 7, no. 3, pp. 153–162, 2003.
1. How often do you do this activity? [11] J. A. Cruz and D. S. Wishart, “Applications of machine learning in
• Everyday (50.3%) • Twice a day (4.5%) cancer prediction and prognosis,” Cancer Informatics, vol. 2, pp. 59–
• Once a week (24%) • Twice a week (21.2%) 78, 2006.
2. Do you feel an urge to do it? [12] K. Kourou, T. P. Exarchos, K. P. Exarchos, M. V. Karamouzis, and
D. I. Fotiadis, “Machine learning applications in cancer prognosis
• Yes (59.2%) • No (8.6%)
and prediction,” Computational and Structural Biotechnology Journal,
• Not really (32.2%)
vol. 13, pp. 8–17, 2015.
3. Does it affect your mood if you do not do this activity on [13] R. Salgado, C. Denkert, S. Demaria, N. Sirtaine, F. Klauschen,
a regular basis? G. Pruneri, S. Wienert, G. Van den Eynden, F. L. Baehner, F. Penault-
• Yes (43.2%) • No (42.8%) Llorca, et al., “The evaluation of tumor-infiltrating lymphocytes (TILs)
• Not sure (14%) in breast cancer: recommendations by an international TILs working
4. Do you do this activity alone or with some company? group 2014,” Annals of Oncology, vol. 26, no. 2, pp. 259–271, 2015.
[14] H. R. Ali, A. Dariush, E. Provenzano, H. Bardwell, J. E. Abraham,
• Alone (50.3%) • Two to three (23.3%) M. Iddawela, A.-L. Vallier, L. Hiller, J. A. Dunn, S. J. Bowden, et al.,
• More than three (3.1%) • Does not matter (23.3%) “Computational pathology of pre-treatment biopsies identifies lympho-
5. Besides work or studies, how many other main activities you have cyte density as a predictor of response to neoadjuvant chemotherapy
in a day besides this activity? in breast cancer,” Breast Cancer Research, vol. 18, no. 1, p. 21, 2016.
• Three more (66.1%) • Five more (14%) [15] M. A. Myszczynska, P. N. Ojamies, A. M. Lacoste, D. Neil, A. Saffari,
• 10 more (1.4%) • Many (18.5%) R. Mead, G. M. Hautbergue, J. D. Holbrook, and L. Ferraiuolo,
“Applications of machine learning to diagnosis and treatment of
6. How many very close friends and family do you talk to everyday? neurodegenerative diseases,” Nature Reviews Neurology, vol. 16, no. 8,
• One or two (44.5%) • Three to five (36%) pp. 440–456, 2020.
• Around 10 (8.2%) • Many (11.3%) [16] I. El-Naqa, Y. Yang, N. P. Galatsanos, R. M. Nishikawa, and M. N.
7. Did you attempt to seek professional help to stop this activity? Wernick, “A similarity learning approach to content-based image
• No (74.7%) • Yes (6.8%) retrieval: application to digital mammography,” Medical Imaging, IEEE
• Not really (18.5%) Transactions on, vol. 23, no. 10, pp. 1233–1244, 2004.
[17] H. Müller, N. Michoux, D. Bandon, and A. Geissbuhler, “A review of
8. Do you think you get distracted in your daily work or studies by content-based image retrieval systems in medical applications—clinical
thinking about this activity? benefits and future directions,” International Journal of Medical Infor-
• Yes (25.7%) • No (48.3%) matics, vol. 73, no. 1, pp. 1–23, 2004.
• Manageable (26%) [18] D. Salas-Gonzalez, J. M. Górriz, J. Ramírez, M. López, I. A. Illan,
9. Have you talked with others about this activity? F. Segovia, C. G. Puntonet, and M. Gómez-Río, “Analysis of spect
• No, I keep it to myself (36.3%) brain images for the diagnosis of alzheimer’s disease using moments
• Selected few (32.2%) and support vector machines,” Neuroscience Letters, vol. 461, no. 1,
pp. 60–64, 2009.
• Close friends and family only (15.4%)
• Everybody knows about it (16.1%) [19] R. Chaves, J. Ramírez, J. Górriz, M. López, D. Salas-Gonzalez,
I. Alvarez, and F. Segovia, “SVM-based computer-aided diagnosis
10. Rate yourself with regard to this activity of the Alzheimer’s disease using t-test NMSE feature selection with
• Addicted (19.5%) • Not addicted (23.3%) feature correlation weighting,” Neuroscience Letters, vol. 461, no. 3,
• Manageable (44.9%) • I don’t know (12.3%) pp. 293–297, 2009.
[20] H. Greenspan, B. van Ginneken, and R. M. Summers, “Guest editorial
deep learning in medical imaging: Overview and future promise of
an exciting new technique,” IEEE Transactions on Medical Imaging,
[3] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik, “Gene selection vol. 35, no. 5, pp. 1153–1159, 2016.
for cancer classification using support vector machines,” Machine [21] L. Macyszyn, H. Akbari, J. M. Pisapia, X. Da, M. Attiah, V. Pigrish,
Learning, vol. 46, no. 1-3, pp. 389–422, 2002. Y. Bi, S. Pal, R. V. Davuluri, L. Roccograndi, et al., “Imaging patterns
[4] Q.-H. Ye, L.-X. Qin, M. Forgues, P. He, J. W. Kim, A. C. Peng, R. Si- predict patient survival and molecular subtype in glioblastoma via
mon, Y. Li, A. I. Robles, Y. Chen, et al., “Predicting hepatitis B virus– machine learning techniques,” Neuro-oncology, vol. 18, no. 3, pp. 417–
positive metastatic hepatocellular carcinomas using gene expression 425, 2016.
profiling and supervised machine learning,” Nature Medicine, vol. 9, [22] M. J. Willemink, W. A. Koszek, C. Hardell, J. Wu, D. Fleischmann,
no. 4, pp. 416–423, 2003. H. Harvey, L. R. Folio, R. M. Summers, D. L. Rubin, and M. P.
[5] R. Shaik and W. Ramakrishna, “Machine learning approaches dis- Lungren, “Preparing medical imaging data for machine learning,”
tinguish multiple stress conditions using stress-responsive genes and Radiology, vol. 295, no. 1, pp. 4–15, 2020.
identify candidate genes for broad resistance in rice,” Plant Physiology, [23] I. Kononenko, “Machine learning for medical diagnosis: history, state
vol. 164, no. 1, pp. 481–495, 2014. of the art and perspective,” Artificial Intelligence in Medicine, vol. 23,
[6] W.-L. Yang, R. D. Kouyos, J. Böni, S. Yerly, T. Klimkait, V. Aubert, no. 1, pp. 89–109, 2001.
A. U. Scherrer, M. Shilaih, T. Hinkley, C. Petropoulos, et al., “Persis- [24] M. N. H. Barakat and A. P. Bradley, “Intelligible support vector
tence of transmitted HIV-1 drug resistance mutations associated with machines for diagnosis of diabetes mellitus,” Information Technology
fitness costs and viral genetic backgrounds,” PLoS Pathogens, vol. 11, in Biomedicine, IEEE Transactions on, vol. 14, no. 4, pp. 1114–1120,
no. 3, p. e1004722, 2015. 2010.
[7] R. Magar, P. Yadav, and A. B. Farimani, “Potential neutralizing [25] L. O’Dwyer, F. Lamberton, A. L. Bokde, M. Ewers, Y. O. Faluyi,
antibodies discovered for novel corona virus using machine learning,” C. Tanner, B. Mazoyer, D. O’Neill, M. Bartley, D. R. Collins, et al.,
Scientific Reports, vol. 11, no. 1, pp. 1–11, 2021. “Using support vector machines with multiple indices of diffusion
[8] O. Shahid, M. Nasajpour, S. Pouriyeh, R. M. Parizi, M. Han, M. Valero, for automated classification of mild cognitive impairment,” PloS One,
F. Li, M. Aledhari, and Q. Z. Sheng, “Machine learning research vol. 7, no. 2, p. e32441, 2012.
towards combating covid-19: Virus detection, spread prevention, and [26] V. Prasad, T. S. Rao, and M. S. P. Babu, “Thyroid disease diagnosis
medical assistance,” Journal of Biomedical Informatics, vol. 117, p. via hybrid architecture composing rough data sets theory and machine
103751, 2021. learning algorithms,” Soft Computing, vol. 20, no. 3, pp. 1179–1189,
[9] S. Dreiseitl, L. Ohno-Machado, H. Kittler, S. Vinterbo, H. Billhardt, 2016.
and M. Binder, “A comparison of machine learning methods for the di- [27] J. G. Richens, C. M. Lee, and S. Johri, “Improving the accuracy of
agnosis of pigmented skin lesions,” Journal of Biomedical Informatics, medical diagnosis with causal machine learning,” Nature Communica-
vol. 34, no. 1, pp. 28–36, 2001. tions, vol. 11, no. 1, pp. 1–9, 2020.
[10] M. Gletsos, S. G. Mougiakakou, G. K. Matsopoulos, K. S. Nikita, [28] W. Zheng, “Multichannel eeg-based emotion recognition via group
A. S. Nikita, and D. Kelekis, “A computer-aided diagnostic system sparse canonical correlation analysis,” IEEE Transactions on Cognitive
to characterize CT focal liver lesions: design and optimization of and Developmental Systems, vol. 9, no. 3, pp. 281–290, 2017.
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION
[29] S. Chen, S. Zhang, J. Shang, B. Chen, and N. Zheng, “Brain-inspired
cognitive model with attention for self-driving cars,” IEEE Transactions
on Cognitive and Developmental Systems, vol. 11, no. 1, pp. 13–25,
2017.
[30] R. Hortensius, F. Hekele, and E. S. Cross, “The perception of emotion
in artificial agents,” IEEE Transactions on Cognitive and Developmen-
tal Systems, vol. 10, no. 4, pp. 852–864, 2018.
[31] L. Jamone, E. Ugur, A. Cangelosi, L. Fadiga, A. Bernardino, J. Piater,
and J. Santos-Victor, “Affordances in psychology, neuroscience, and
robotics: A survey,” IEEE Transactions on Cognitive and Develop-
mental Systems, vol. 10, no. 1, pp. 4–25, 2018.
[32] B. A. Cociu, S. Das, L. Billeci, W. Jamal, K. Maharatna, S. Calderoni,
A. Narzisi, and F. Muratori, “Multimodal functional and structural brain
connectivity analysis in autism: A preliminary integrated approach with
eeg, fmri, and dti,” IEEE Transactions on Cognitive and Developmental
Systems, vol. 10, no. 2, pp. 213–226, 2018.
[33] T. N. Malete, K. Moruti, T. S. Thapelo, and R. S. Jamisola, “Eeg-
based control of a 3d game using 14-channel emotiv epoc+,” in 2019
IEEE International Conference on Cybernetics and Intelligent Systems
(CIS) and IEEE Conference on Robotics, Automation and Mechatronics
(RAM), 2019, pp. 463–468.
[34] W. Mmereki, R. S. Jamisola, D. Mpoeleng, and T. Petso, “Yolov3-based
human activity recognition as viewed from a moving high-altitude
aerial camera,” in 2021 7th International Conference on Automation,
Robotics and Applications (ICARA), 2021, pp. 241–246.
[35] L. O. Mohutsiwa and R. S. Jamisola, “Eeg-based human emotion
classification using combined computational techniques for feature
extraction and selection in six machine learning models,” in 2021
5th International Conference on Intelligent Computing and Control
Systems (ICICCS), 2021, pp. 1095–1102.
[36] R. W. Picard, E. Vyzas, and J. Healey, “Toward machine emotional
intelligence: Analysis of affective physiological state,” Pattern Analysis
and Machine Intelligence, IEEE Transactions on, vol. 23, no. 10, pp.
1175–1191, 2001.
[37] A. Ben-David and J. Mandel, “Classification accuracy: Machine learn-
ing vs. explicit knowledge acquisition,” Machine Learning, vol. 18,
no. 1, pp. 109–114, 1995.
[38] J. Bollen, H. Mao, and A. Pepe, “Modeling public mood and emotion:
Twitter sentiment and socio-economic phenomena,” in Proceedings of
the International AAAI Conference on Web and Social Media, vol. 5,
no. 1, 2011.
[39] R. S. Jamisola Jr, E. P. Dadios, and M. H. Ang Jr, “Using metaheuristic
computations to find the minimum-norm-residual solution to linear
systems of equations,” Philippine Computing Journal, vol. 4, no. 2,
pp. 1–9, 2009.
[40] R. S. Jamisola Jr, “Of love and affection and the gaze sensor,” Lovotics,
vol. 1, no. 1, p. 10000e102, 2014.
[41] Z. Ghahramani, “Probabilistic machine learning and artificial intelli-
gence,” Nature, vol. 521, no. 7553, pp. 452–459, 2015.
[42] A. Holzinger, “Interactive machine learning for health informatics:
when do we need the human-in-the-loop?” Brain Informatics, vol. 3,
no. 2, pp. 119–131, 2016.
[43] G. Adomavicius and A. Tuzhilin, “Context-aware recommender sys-
tems,” in Recommender Systems Handbook. Springer, 2015, pp. 191–
226.
[44] J. M. Hofman, A. Sharma, and D. J. Watts, “Prediction and explanation
in social systems,” Science, vol. 355, no. 6324, pp. 486–488, 2017.
[45] G. Christo, S. L. Jones, S. Haylett, G. M. Stephenson, R. M. Lefever,
and R. Lefever, “The shorter promis questionnaire: Further validation
of a tool for simultaneous assessment of multiple addictive behaviours,”
Addictive Behaviors, vol. 28, no. 2, pp. 225–248, 2003.
[46] Q. J. Huys, T. V. Maia, and M. J. Frank, “Computational psychiatry
as a bridge from neuroscience to clinical applications,” Nature Neuro-
science, vol. 19, no. 3, p. 404, 2016.
[47] J. Bi, J. Sun, Y. Wu, H. Tennen, and S. Armeli, “A machine learning
approach to college drinking prediction and risk factor identification,”
ACM Transactions on Intelligent Systems and Technology (TIST),
vol. 4, no. 4, p. 72, 2013.
[48] A. Markowetz, K. Błaszkiewicz, C. Montag, C. Switala, and T. E.
Schlaepfer, “Psycho-informatics: big data shaping modern psychomet-
rics,” Medical Hypotheses, vol. 82, no. 4, pp. 405–411, 2014.
[49] O. Chapelle and Z. Harchaoui, “A machine learning approach to
conjoint analysis,” Advances in Neural Information Processing Systems,
vol. 17, pp. 257–264, 2005.
[50] S. Maldonado, R. Montoya, and J. López, “Embedded heterogeneous
feature selection for conjoint analysis: A svm approach using l1
penalty,” Applied Intelligence, vol. 46, no. 4, pp. 775–787, 2017.
13
[51] D. Huang and L. Luo, “Consumer preference elicitation of complex
products using fuzzy support vector machine active learning,” Market-
ing Science, vol. 35, no. 3, pp. 445–464, 2016.
[52] E. Aïmeur, G. Brassard, H. Dufort, and S. Gambs, “Clarisse: A machine
learning tool to initialize student models,” in Intelligent Tutoring
Systems. Springer, 2002, pp. 718–728.
[53] G. Barata, S. Gama, J. Jorge, and D. Gonçalves, “Early prediction of
student profiles based on performance and gaming preferences,” IEEE
Transactions on Learning Technologies, vol. 9, no. 3, pp. 272–284,
2016.
[54] K. Yamanishi and H. Li, “Mining open answers in questionnaire data,”
Intelligent Systems, IEEE, vol. 17, no. 5, pp. 58–63, 2002.
[55] N. B. Kersting, B. L. Sherin, and J. W. Stigler, “Automated scoring
of teachers’ open-ended responses to video prompts bringing the
classroom-video-analysis assessment to scale,” Educational and Psy-
chological Measurement, vol. 74, no. 6, pp. 950–974, 2014.
[56] D. C. Wheeler, I. Burstyn, R. Vermeulen, K. Yu, S. M. Shortreed,
A. Pronk, P. A. Stewart, J. S. Colt, D. Baris, M. R. Karagas, et al.,
“Inside the black box: starting to uncover the underlying decision
rules used in a one-by-one expert assessment of occupational exposure
in case-control studies,” Occupational and Environmental Medicine,
vol. 70, no. 3, pp. 203–210, 2013.
[57] T. Terano and Y. Ishino, “Knowledge acquisition from questionnaire
data using simulated breeding and inductive learning methods,” Expert
Systems with Applications, vol. 11, no. 4, pp. 507–518, 1996.
[58] K.-I. Karstoft, A. Statnikov, S. B. Andersen, T. Madsen, and I. R.
Galatzer-Levy, “Early identification of posttraumatic stress following
military deployment: application of machine learning methods to a
prospective study of danish soldiers,” Journal of Affective Disorders,
vol. 184, pp. 170–175, 2015.
[59] G. Z. Oztas and S. Erdem, “Framework selection for developing
optimization algorithms: assessing preferences by conjoint analysis and
best–worst method,” Soft Computing, vol. 25, no. 5, pp. 3831–3848,
2021.
[60] A. H. Wani, A. E. Aiello, G. S. Kim, F. Xue, C. L. Martin,
A. Ratanatharathorn, A. Qu, K. Koenen, S. Galea, D. E. Wildman,
et al., “The impact of psychopathology, social adversity and stress-
relevant dna methylation on prospective risk for post-traumatic stress:
a machine learning approach,” Journal of Affective Disorders, vol. 282,
pp. 894–905, 2021.
[61] Q. Ye, Z. Zhang, and R. Law, “Sentiment classification of online
reviews to travel destinations by supervised machine learning ap-
proaches,” Expert Systems with Applications, vol. 36, no. 3, Part 2,
pp. 6527 – 6535, 2009.
[62] A. Ortigosa, J. M. Martín, and R. M. Carro, “Sentiment analysis
in facebook and its application to e-learning,” Computers in Human
Behavior, vol. 31, pp. 527–541, 2014.
[63] A. Smith and A. Blandford, “Mltutor: An application of machine
learning algorithms for an adaptive web-based information system,”
International Journal of Artificial Intelligence in Education, vol. 13,
no. 2-4, pp. 235–261, 2003.
[64] S. Bidel, L. Lemoine, F. Piat, T. Artieres, and P. Gallinari, “Statistical
machine learning for tracking hypermedia user behavior,” in 2nd Work-
shop on Machine Learning, Information Retrieval and User Modeling,
9th Int. Conf. in UM, 2003.
[65] Y.-L. Zhao, L. Nie, X. Wang, and T.-S. Chua, “Personalized recom-
mendations of locally interesting venues to tourists via cross-region
community matching,” ACM Transactions on Intelligent Systems and
Technology (TIST), vol. 5, no. 3, p. 50, 2014.
[66] F. Razzaghi, H. Minaee, and A. A. Ghorbani, “Context free frequently
asked questions detection using machine learning techniques,” in Web
Intelligence (WI), 2016 IEEE/WIC/ACM International Conference on.
IEEE, 2016, pp. 558–561.
[67] S. Damani, K. N. Narahari, A. Chatterjee, M. Gupta, and P. Agrawal,
“Optimized transformer models for faq answering,” Advances in
Knowledge Discovery and Data Mining, vol. 12084, p. 235, 2020.
[68] L. A. Millard, P. A. Flach, and J. P. Higgins, “Machine learning to assist
risk-of-bias assessments in systematic reviews,” International Journal
of Epidemiology, vol. 45, no. 1, pp. 266–277, 2016.
[69] W. Didimo, L. Grilli, G. Liotta, L. Menconi, F. Montecchiani, and
D. Pagliuca, “Combining network visualization and data mining for
tax risk assessment,” IEEE Access, vol. 8, pp. 16 073–16 086, 2020.
[70] F. N. Koutanaei, H. Sajedi, and M. Khanbabaei, “A hybrid data
mining model of feature selection algorithms and ensemble learning
classifiers for credit scoring,” Journal of Retailing and Consumer
Services, vol. 27, pp. 11–23, 2015.
14
[71] S. Moro, P. Cortez, and P. Rita, “A data-driven approach to predict the
success of bank telemarketing,” Decision Support Systems, vol. 62, pp.
22–31, 2014.
[72] H. Zacharatos, C. Gatzoulis, and Y. L. Chrysanthou, “Automatic
emotion recognition based on body movement analysis: a survey,” IEEE
Computer Graphics and Applications, vol. 34, no. 6, pp. 35–45, 2014.
[73] R. Lotfian and C. Busso, “Practical considerations on the use of
preference learning for ranking emotional speech,” in Acoustics, Speech
and Signal Processing (ICASSP), 2016 IEEE International Conference
on. IEEE, 2016, pp. 5205–5209.
[74] A. L. Chun, G. J. Anderson, and A. Yosher, “Determining and
communicating user’s emotional state related to user’s physiological
and non-physiological data,” Aug. 16 2016, US Patent 9,418,390.
[75] P. Rani, C. Liu, N. Sarkar, and E. Vanman, “An empirical study of
machine learning techniques for affect recognition in human–robot
interaction,” Pattern Analysis and Applications, vol. 9, no. 1, pp. 58–69,
2006.
[76] P. Chalfoun, S. Chaffar, and C. Frasson, “Predicting the emotional
reaction of the learner with a machine learning technique,” in Workshop
on Motivaional and Affective Issues in ITS, ITS’06, International
Conference on Intelligent Tutoring Systems. Citeseer, 2006.
[77] H. Al-Samarraie, S. M. Sarsam, and H. Guesgen, “Predicting user
preferences of environment design: a perceptual mechanism of user
interface customisation,” Behaviour & Information Technology, vol. 35,
no. 8, pp. 644–653, 2016.
[78] A. Bajoulvand, R. Z. Marandi, M. R. Daliri, and S. H. Sabzpoushan,
“Analysis of folk music preference of people from different ethnic
groups using kernel-based methods on eeg signals,” Applied Mathe-
matics and Computation, vol. 307, pp. 62–70, 2017.
[79] L. H. Chew, J. Teo, and J. Mountstephens, “Aesthetic preference
recognition of 3d shapes using eeg,” Cognitive Neurodynamics, vol. 10,
no. 2, pp. 165–173, 2016.
[80] H. J. Cha, Y. S. Kim, S. H. Park, T. B. Yoon, Y. M. Jung, and J.-
H. Lee, “Learning styles diagnosis based on user interface behaviors
for the customization of learning interfaces in an intelligent tutoring
system,” in Intelligent Tutoring Systems. Springer, 2006, pp. 513–524.
[81] S. Stumpf, V. Rajaram, L. Li, W.-K. Wong, M. Burnett, T. Dietterich,
E. Sullivan, and J. Herlocker, “Interacting meaningfully with machine
learning systems: Three experiments,” International Journal of Human-
Computer Studies, vol. 67, no. 8, pp. 639–662, 2009.
[82] A. Abid, A. Abdalla, A. Abid, D. Khan, A. Alfozan, and J. Zou,
“An online platform for interactive feedback in biomedical machine
learning,” Nature Machine Intelligence, vol. 2, no. 2, pp. 86–88, 2020.
[83] M. S. Bartlett, G. Littlewort, M. Frank, C. Lainscsek, I. Fasel, and
J. Movellan, “Recognizing facial expression: machine learning and
application to spontaneous behavior,” in Computer Vision and Pattern
Recognition, 2005. CVPR 2005. IEEE Computer Society Conference
on, vol. 2. IEEE, 2005, pp. 568–573.
[84] G. Littlewort, M. S. Bartlett, I. Fasel, J. Susskind, and J. Movellan,
“Dynamics of facial expression extracted automatically from video,”
Image and Vision Computing, vol. 24, no. 6, pp. 615–625, 2006.
[85] N. Sebe, M. S. Lew, Y. Sun, I. Cohen, T. Gevers, and T. S. Huang,
“Authentic facial expression analysis,” Image and Vision Computing,
vol. 25, no. 12, pp. 1856–1863, 2007.
[86] C. Shan, S. Gong, and P. W. McOwan, “Facial expression recognition
based on local binary patterns: A comprehensive study,” Image and
Vision Computing, vol. 27, no. 6, pp. 803–816, 2009.
[87] S. V. Ioannou, A. T. Raouzaiou, V. A. Tzouvaras, T. P. Mailis, K. C.
Karpouzis, and S. D. Kollias, “Emotion recognition through facial
expression analysis based on a neurofuzzy network,” Neural Networks,
vol. 18, no. 4, pp. 423–435, 2005, Special Issue: Emotion and Brain.
[88] I. Cohen, N. Sebe, A. Garg, L. S. Chen, and T. S. Huang, “Facial
expression recognition from video sequences: temporal and static
modeling,” Computer Vision and Image Understanding, vol. 91, no. 1,
pp. 160–187, 2003.
[89] L. Yin, X. Wei, Y. Sun, J. Wang, and M. J. Rosato, “A 3d facial ex-
pression database for facial behavior research,” in Automatic Face and
Gesture Recognition, 2006. FGR 2006. 7th International Conference
on. IEEE, 2006, pp. 211–216.
[90] S. Lin, M. Bai, F. Liu, L. Shen, and Y. Zhou, “Orthogonalization-
guided feature fusion network for multimodal 2d+ 3d facial expression
recognition,” IEEE Transactions on Multimedia, vol. 23, pp. 1581–
1591, 2020.
[91] C. A. Wall, Q. Wang, M. Weng, E. S. Kim, L. Whitaker, M. Perlmut-
ter, and F. Shic, “Mapping connections between biological-emotional
preferences and affective recognition: An eye-tracking interface for
passive assessment of emotional competency,” in Affective Computing
ARXIV.ORG » COMPUTER SCIENCE » MACHINE LEARNING
and Intelligent Interaction (ACII), 2015 International Conference on.
IEEE, 2015, pp. 21–27.
[92] G. C.-L. Hung, P.-C. Yang, C.-C. Chang, J.-H. Chiang, and Y.-Y. Chen,
“Predicting negative emotions based on mobile phone usage patterns:
an exploratory study,” JMIR Research Protocols, vol. 5, no. 3, p. e160,
2016.
[93] O. Toubia, T. Evgeniou, and J. Hauser, “Optimization-based and
machine-learning methods for conjoint analysis: Estimation and ques-
tion design,” in Conjoint Measurement: Methods and Applications,
4th ed., A. Gustafsson, A. Herrmann, and F. Huber, Eds. New York,
NY: Springer, 2007, ch. 12, pp. 231–258.
[94] X. Li, H. Li, B. Pan, and R. Law, “Machine learning in internet search
query selection for tourism forecasting,” Journal of Travel Research,
vol. 60, no. 6, pp. 1213–1231, 2021.
[95] E. Hüllermeier, J. Fürnkranz, W. Cheng, and K. Brinker, “Label ranking
by learning pairwise preferences,” Artificial Intelligence, vol. 172, no.
16–17, pp. 1897 – 1916, 2008.
[96] G. S. Budhi, R. Chiong, I. Pranata, and Z. Hu, “Using machine
learning to predict the sentiment of online reviews: a new framework
for comparative analysis,” Archives of Computational Methods in
Engineering, vol. 28, no. 4, pp. 2543–2566, 2021.
[97] H. Kumar, A. Solanki, K. K. Singh, et al., “Progressive machine
learning approach with webastro for web usage mining,” Procedia
Computer Science, vol. 167, pp. 1400–1410, 2020.
[98] L. Devillers, L. Vidrascu, and L. Lamel, “Challenges in real-life
emotion annotation and machine learning based detection,” Neural
Networks, vol. 18, no. 4, pp. 407 – 422, 2005, Special Issue: Emotion
and Brain.
[99] D. Freitag, “Machine learning for information extraction in informal
domains,” Machine Learning, vol. 39, no. 2-3, pp. 169–202, 2000.
[100] A. Agrawal and A. Jain, “Speech emotion recognition of hindi speech
using statistical and machine learning techniques,” Journal of Interdis-
ciplinary Mathematics, vol. 23, no. 1, pp. 311–319, 2020.
[101] A. Fathima and K. Vaidehi, “Review on facial expression recognition
system using machine learning techniques,” in Advances in Decision
Sciences, Image Processing, Security and Computer Vision. Springer,
2020, pp. 608–618.
[102] B. Jin, Y. Qu, L. Zhang, and Z. Gao, “Diagnosing parkinson disease
through facial expression recognition: video analysis,” Journal of
Medical Internet Research, vol. 22, no. 7, p. e18697, 2020.
[103] J. E. Beck and B. P. Woolf, “High-level student modeling with machine
learning,” in Intelligent Tutoring Systems. Springer, 2000, pp. 584–
593.
[104] R. Lamb, B. Hand, and A. Kavner, “Computational modeling of the
effects of the science writing heuristic on student critical thinking in
science using machine learning,” Journal of Science Education and
Technology, vol. 30, no. 2, pp. 283–297, 2021.
[105] S. Cetintas, L. Si, Y. P. Xin, and C. Hord, “Automatic detection of off-
task behaviors in intelligent tutoring systems with machine learning
techniques,” Learning Technologies, IEEE Transactions on, vol. 3,
no. 3, pp. 228–236, 2010.
[106] Z. E. Abou Elassad, H. Mousannif, H. Al Moatassime, and A. Kark-
ouch, “The application of machine learning techniques for driving
behavior analysis: A conceptual framework and a systematic literature
review,” Engineering Applications of Artificial Intelligence, vol. 87, p.
103312, 2020.
[107] G. I. Webb, M. J. Pazzani, and D. Billsus, “Machine learning for user
modeling,” User Modeling and User-Adapted Interaction, vol. 11, no.
1-2, pp. 19–29, 2001.
[108] J. A. Walonoski and N. T. Heffernan, “Detection and analysis of off-
task gaming behavior in intelligent tutoring systems,” in Intelligent
Tutoring Systems. Springer, 2006, pp. 382–391.
[109] E. Rolf, M. Simchowitz, S. Dean, L. T. Liu, D. Bjorkegren, M. Hardt,
and J. Blumenstock, “Balancing competing objectives with noisy
data: Score-based classifiers for welfare-aware machine learning,” in
International Conference on Machine Learning. PMLR, 2020, pp.
8158–8168.
[110] J. H. Boose, “A knowledge acquisition program for expert systems
based on personal construct psychology,” International Journal of Man-
Machine Studies, vol. 23, no. 5, pp. 495 – 525, 1985.
[111] J. B. Tenenbaum, T. L. Griffiths, and C. Kemp, “Theory-based bayesian
models of inductive learning and reasoning,” Trends in Cognitive Sci-
ences, vol. 10, no. 7, pp. 309 – 318, 2006, Special Issue: Probabilistic
Models of Cognition.
[112] Y. Huang, Z. Fu, and C. L. Franzke, “Detecting causality from time
series in a machine learning framework,” Chaos: An Interdisciplinary
Journal of Nonlinear Science, vol. 30, no. 6, p. 063116, 2020.
JAMISOLA: A MACHINE LEARNING TOOL TO DETERMINE STATE OF MIND AND EMOTION
[113] M. P. Jarvis, G. Nuzzo-Jones, and N. T. Heffernan, “Applying machine
learning techniques to rule generation in intelligent tutoring systems,”
in Intelligent Tutoring Systems. Springer, 2004, pp. 541–553.
[114] G. Castillo, J. Gama, and A. M. Breda, “Adaptive bayes for a student
modeling prediction task based on learning styles,” in User Modeling
2003. Springer, 2003, pp. 328–332.
[115] V. Vapnik, Statistical Learning Theory. Wiley, New York, 1998.
15
[116] R. Jamisola, “A Survey on Addiction,” https://docs.google.com/forms/d/
1iN779Pe0PCo8ejybrG2dOlrJ0fG-JYMXTRA6PWME5V4/viewform,
2016, [Online; accessed 26-May-2020].
[117] D. Raistrick, G. Dunbar, and R. Davidson, “Development of a ques-
tionnaire to measure alcohol dependence,” British Journal of Addiction,
vol. 78, no. 1, pp. 89–95, 1983.