SAPaper ML

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Copyright © 2018 American Scientific Publishers Journal of

All rights reserved Computational and Theoretical Nanoscience


Printed in the United States of America Vol. 15, 1662–1665, 2018

A Survey Based on Human Emotion Identification Using


Machine Learning and Deep Learning
C. Akalya Devi1 , D. KarthikaRenuka1 , and S. Soundarya2 ∗
1
Dept. of Information Technology, PSG College of Technology, Coimbatore, Tamil Nadu, India
2
PG Scholar, Dept. of Information Technology, PSG College of Technology, Coimbatore, Tamil Nadu, India

An Emotion can be identified by various activities of a person. A human can express their feelings or
RESEARCH ARTICLE

Emotions in different ways. Emotions are classified into different types like happy, disgust, anger and
so on. Emotions can be detected or identified in variety of ways. Emotion can be recognized by body
gestures, EEG signals, face images, text and so on. A survey or review is made on identifying the
human emotions using EEG signals with help of Machine Learning and Deep Learning Techniques.
In this paper, a review is done to find the best algorithms for finding an Emotion of a person by
analyzing various Machine Learning and Deep Learning algorithms with the help of previous papers.
Keywords: Emotions, Machine Learning, Deep Learning, EEG.

1. INTRODUCTION 2. Another more specific that considers the aspects


Emotion plays major role in human IP: 45.148.233.50 onlyMon,involved
daily life, not On: 27 Apr in the process
2020 21:47:03of recognizing emotions from EEG
for interaction but also in many part Copyright:
of our life. American
Emotion Scientific Publishers
signals (e.g., subjects, features extracted, classifiers, etc.).
detection technology is making a huge difference Delivered
in how byForIngenta
this they analysed all papers from 2009 to 2016 and
we leverage text analysis and EEG signals. Especially, in they go through each and every algorithm they used for
the field of marketing and health care. Emotion Detec- EEG signals pre-processing, feature extraction and clas-
tion and Recognition from EEG signals is a recent field of sification. In this survey paper, they derive a set of best
research that is closely related to Sentiment Analysis. Sen- practice recommendations to help researchers produce well
timent Analysis aims to detect positive, neutral, or nega- validated and high-quality works, and also able to be
tive feelings from EEG signals, whereas Emotion Analysis reproducible and replicable.
aims to detect and recognize types of feelings through the In Paper [2], they reviewed emotion classifications,
expression of texts, such as anger, disgust, fear, happiness, emotion evoking experiments and emotion recognition
sadness, and surprise. A survey on various techniques in algorithms. They proposed and implemented a novel frac-
machine learning and deep learning for detecting or iden- tal dimension based algorithm for recognition of emotions
tifying human emotions has been discussed in this paper. from EEG in real time. They implemented the algorithm
along with Haptek system.
2. RELATED WORKS In Paper [3], a DREAMER database had taken for anal-
Emotion detection using EEG signals is a universal issue ysis of emotion from EEG signals. The database includes
and causes difficulties due to uncertain physical and psy- EEG and ECG recordings from 23 participants, where each
chological characteristics of emotions. participant rated his emotional response along the scales
In Paper [1], they analysed the previous works from of valence, arousal, and dominance, after watching each of
2009 to 2016, that propose novel methods for the recog- the 18 film clips selected to elicit specific emotions. The
nition of emotions through EEG signals. Their analysis DREAMER database made publically available after pub-
draws on two perspectives: lication of this paper. In order to give the opportunity to
1. A set of recommendations to avoid the common pitfalls researchers to evaluate their algorithm using this database.
of this type of research area. In this Paper [4], they classified emotions using EEG
signals using various Deep Learning and Machine Learn-
ing algorithms. They compared the results each and every

Author to whom correspondence should be addressed. algorithm they used for emotion classification. In their

1662 J. Comput. Theor. Nanosci. 2018, Vol. 15, No. 5 1546-1955/2018/15/1662/004 doi:10.1166/jctn.2018.7358
Devi et al. A Survey Based on Human Emotion Identification Using Machine Learning and Deep Learning

work they used Deep models like DBN (deep belif net- above. They conclude from that neural network and SVM
works) and DBN-HMM (A hidden markov model (HMM) provides higher accuracy when compared with other algo-
is integrated along with DBN). They also compare the rithm mentioned above.
performance of the deep models to KNN, SVM and In Paper [8], they made a survey on classification and
Graph regularized Extreme Learning Machine (GELM). feature extraction for EEG signals. The feature extrac-
The average accuracies they got for DBN-HMM, DBN, tion techniques they used for their study, were PCA and
GELM, SVM, and KNN in their experiments are 87.62%, ICA techniques. The classification algorithms they used
86.91%, 85.67%, 84.08%, and 69.66%, respectively. From for their study were Support Vector Machine and K-means
their experimental results they show that DBN and DBN- Algorithm. They just discussed about the methods in their
HMM models improve the accuracy of EEG based emo- survey paper.
tion classification in comparison with other methods. In Paper [9], they done a survey based on feature extrac-
In Paper [5], they used HCNN (hierarchical convolu- tion methods and feature classification techniques. They
tional neural network) for classification of different states discussed the advantages and disadvantages of each and
of emotion. They also used stacked autoencoder (SAE), every method they mentioned in their paper for feature
SVM, and KNN as competing methods. As a result extraction and feature classification techniques for classi-
deep models (HCNN and SAE) show absolute advantage fication of different types of emotions using EEG signals.

RESEARCH ARTICLE
over traditional shallow models (SVM and KNN). They In Paper [10], the discussion done to tell the effective-
confirmed that the high-frequency wave bands Beta and ness of Neural Networks to classify user emotions using
Gamma were the most suitable bands for emotion read- EEG signals from the DEAP database. They used two
ing. HCNN gives them the higher accuracies when com- deep learning models to classify the user emotion states.
pared with other methods, and SAE is slightly inferior than The deep learning models they used were DNN and CNN.
HCNN. They classified the emotion for valance and arousal classes
In Article [6], the classification of emotion have imple- using EEG signals from DEAP database. From their com-
mented using the DLN (deep learning network) and parison they found CNN provides better results when com-
machine learning algorithms like SVM and Naive Bayes pared to DNN. They also compare our results with that
classifier. Then they enhanced with covariate shift adapta- of, for classification in 3 classes and for classification in
tion of the principal components. The deep learning net- 2 classes, which represent the previous state of the art
IP: 45.148.233.50 On:
work they used constitute of three stack autoencoders and Mon, 27 Apr 2020
classification 21:47:03
accuracy for both Valence and Arousal.
Copyright: American
two softmax classifiers for valence and arousal state clas- Scientific Publishers
Delivered by Ingenta
sifications. As a result of their work, they got classifica- 3. MACHINE LEARNING TECHNIQUES
tion accuracy of the DLN with PCA + CSA is 53.42% and
Machine learning is a scientific discipline that deals with
52.05% to classify three levels of valence states and three
the construction and study of algorithms that can learn
levels of arousal states. Thus they conclude DLN provides
from data (R. Kohavi and F. Provost 1998). Such algo-
better accuracy performance compared to SVM and Naive
rithms operate by building a model based on inputs and
Bayes classifier.
using these inputs to make predictions or decisions, rather
In Paper [7], the different feature extraction and classi-
than following only explicitly programmed instructions
fication algorithms for EEG signals have discussed. In this
(C. M. Bishop, 2006). Specifically in emotion detec-
paper, they focused on several feature extraction methods
tion, Machine learning algorithms are used to learn how
like,
to detect emotions. These approaches can be divided
• Discrete wavelet Transformation
into supervised and unsupervised learning. A. Supervised
• Higher Order Crossing
Learning approaches rely on a labelled training data, a set
• Principal Component Analysis
of training examples. The supervised learning algorithm
• Nonlinear Dynamic Analysis
analyses the training data and infers a function, which we
• Independent Component Analysis
use for mapping new examples.
• Short Time Fourier Transform and Mutual
In the Machine Learning (ML) approach to sentiment
• Information
analysis, the classifier automatically learns the properties
• Statistical Based Feature.
and categories from the pre-classified training documents.
The classification algorithms they discussed in this ML based classification is called as supervised learning
paper is, because this process is guided by the labeled training set.
• Neural Network There exists various machine learning approaches, some
• K-Nearest Neighbor of them are explained as below.
• Support vector Machine
• Linear Dynamic Analysis. 3.1. Support Vector Machines
They discussed the advantages and disadvantages of all The fastest and most widely used algorithm among all
feature extraction and classification algorithms mentioned the ML algorithms is the support vector machine (SVM).

J. Comput. Theor. Nanosci. 15, 1662–1665, 2018 1663


A Survey Based on Human Emotion Identification Using Machine Learning and Deep Learning Devi et al.

A binary SVM is a hyperplane separating the feature space 3.5. Neural Networks
of positive instances from the feature space of negative Neural network (NN) can be designed to carry out the
instances. During the training phase, the hyperplane that task of opinion mining. The features of a document are
can separate the positive feature space from the nega- the input nodes, the output nodes deliver the category. The
tive feature space with a maximal margin is chosen. The dependence relations are taken care by the link weights
margin is distance of the nearest point from the posi- (M. Lam, 2004). Generally the NN are trained by back
tive and negative sets to the hyperplane. Support vec- propagation, i.e., the training documents are fed into the
tors, the subset of the training instances, determine the input nodes and if a wrong classification occurs, the error
hyperplane for a SVM. SVM classifiers perform extremely is propagated back in the network to minimize the error
well irrespective of the dimensionality of the feature by adjusting the link weights. Perceptron is the simplest
space. kind of a NN, which has only two layers viz. input layer
and output layer. A multilayer perceptron contains one or
3.2. Decision Tree Classifiers more hidden layers between the input and output layers.
A Decision Tree (DT) classifier resembles a tree in which
the features are represented by nodes and the edges leav- 4. DEEP LEARNING TECHNIQUES
RESEARCH ARTICLE

ing a node are labeled by the feature weight and leaves Emotion Detection can be performed efficiently by imple-
represent the categories. The tree is constructed based on menting different deep learning models, which have been
a recursive procedure. At each step a feature ‘F’ is picked extended recently. These models include CNN (convolu-
and the training collection is divided into two groups, tional neural networks), DNN (deep neural networks) and
one containing ‘F’ and another not containing ‘F’. This DBN (deep belief networks). This describes the efforts of
procedure is carried out until only documents of a single different researchers toward implementing deep learning
category remain. A leaf is generated at the end of this pro- models for performing the emotion detection.
cedure. Information gain or entropy is used to choose a
feature at each step. Generally the DT classifiers are used 4.1. Convolutional Neural Networks (CNN)
as baseline classifiers. The CNN (convolutional neural network) includes pooling
layers and sophistication as it gives a standard architecture
3.3. Fuzzy Logic IP: 45.148.233.50 On: Mon,to27 Aprthe
map 2020 21:47:03
signals of variable length into signals of fixed
Fuzzy logic based systems are handy in Copyright: American Scientific
real life situations Publishers
size scattered vectors. This study has proposed a novel
where the decision to be taken are based on multiple Delivered
crite- byconvolutional
Ingenta
neural network (CNN) framework for clas-
ria with complex interlink among them. It is very true for sification of emotions using EEG signals. CNN has been
a sentiment analysis process in which the system must be implemented using Tensorflow on a windows machine. As
able to understand the sentiment expressed by a customer CNN enhance its performance by increasing its size and
in a review based on the statements about various features depth, so a very deep CNN model, inspired by Tensorflow
of the product or service. For example, in a movie review, is proposed with 22 layers for Emotion analysis. CNN use
the reviewer may praise the director for the usage of tech- four different layers in it for classification. The layers in
nology and blame on the acting and story. Deciding on the CNN were convolutional layer, pooling layer, flattening
overall sentiment as positive or negative depends on the layer and the fully connected layer. Each layer performs
opinion words or phrases used by the reviewer for each some operation and output from each layer will be given
of the features. When the number of features is more, the as a input to the next layer. Pooling layer will perform
complexity in the decision making gets added and hence down-sampling in it. Fully connected layer is the layer
the decision making becomes tough. In such situations, where we get the classification result.
fuzzy logic can be effectively used.
4.2. Deep Belief Networks (DBN)
3.4. Probabilistic Classifier Deep belief networks (DBNs) includes several hidden lay-
Naïve Bayes Algorithm Probabilistic classifiers use the ers, composed by RBM (restricted Boltzmann machines).
Bayes’ theorem to calculate the probability P (c  d), that DBN has been proved efficient for feature representation.
a document belongs to a category c. It utilizes the unlabeled data and fulfills the deficiencies of
labelled analysis issues. The features of language specific
P c  d = P d  cP c/P d and inter language have been presented through building
multiple weakly shared layers of features. In compari-
In order to determine the probability P d  c, it is assumed son with existing studies the proposed work address the
that the coordinates representing the document as a fea- challenge of shortening overlap among feature spaces of
ture vector are independent. These classifiers are called as both source and target language data through cross lin-
Naïve Bayes (NB) classifiers. gual information transfer process using back propagation.

1664 J. Comput. Theor. Nanosci. 15, 1662–1665, 2018


Devi et al. A Survey Based on Human Emotion Identification Using Machine Learning and Deep Learning

DNNs used for transformation of information from source discussed in this review to provide a deep knowledge of
to target language. The experiments have been conducted the successful growing of deep learning applications in the
for emotion classification tasks of EEG signals taken from field of Emotion analysis.
brain waves. When we add more hidden layers it gives
better results for us.
References
1. S. M. Alarcao and M. J. Fonseca, IEEE Transactions on Affective
5. CONCLUSION Computing (2017).
As for Machine Learning approaches, the supervised learn- 2. Y.-J. Liu, M. Yu, G. Zhao, J. Song, Y. Ge, and Y. Shi, IEEE Trans-
actions on Affective Computing (2017).
ing approach is more used in emotion detection because it 3. S. Katsigiannis and Naeem Ramzan, IEEE Journal of Biomedical
usually leads to better results than unsupervised learning. and Health Informatics (2017).
The unsupervised learning approaches need labelling train- 4. W.-L. Zheng, J.-Y. Zhu, Y. Peng, and B.-L. Lu, EEG-based emotion
ing examples and annotating of examples, which is a time- classification using deep belief networks, 2014 IEEE International
consuming task. For this reason, several researches have Conference on Multimedia and Expo (ICME), IEEE (2014), pp. 1–6.
analyzed and realized. Although, it is important to use 5. J. Li, Z. Zhang, and H. He, Cognitive Computation 1 (2017).
6. S. Jirayucharoensak, Setha Pan-Ngum, and Pasin Israsena, The Sci-
good feature extraction techniques in Machine Learning
entific World Journal 2014 (2014).

RESEARCH ARTICLE
algorithms to obtain good results. Moreover, in emotional 7. M. Sreeshakthy, J. Preethi, and A. Dhilipan, International Journal
detection systems based on machine learning approach, of Information Technology and Computer Science (IJITCS) 8, 19
we have detected that most of these systems use features (2016).
based on a pre-processed EEG signals. Hence, we propose 8. Shamla Mantri, Vipul Patil, and Rachana Mitkar, International Jour-
a new direction focuses on deep model analysis, since we nal of Engineering Research and Development 4, 24 (2012).
consider that if we use features based on a deep analysis 9. F. Lotte, M. Congedo, A. Lécuyer, F. Lamarche, and B. Arnaldi,
Journal of Neural Engineering 4, R1 (2007).
on EEG signals we could improve the emotional detection 10. Samarth Tripathi, Shrinivas Acharya, Ranti Dev Sharma, Sudhanshu
systems. Deep learning consists of numerous effective and Mittal, and Samit Bhattacharya, Using Deep and Convolutional Neu-
popular models, these models are used to solve the vari- ral Networks for Accurate Emotion Classification on DEAP Dataset,
ety of problems effectively. Different studies have been AAAI (2017), pp. 4746–4752.

IP: 45.148.233.50 On: Mon, 27 AprReceived: 24 April 2018. Accepted: 9 May 2018.
2020 21:47:03
Copyright: American Scientific Publishers
Delivered by Ingenta

J. Comput. Theor. Nanosci. 15, 1662–1665, 2018 1665

You might also like