An enhanced fitness function to recognize unbalanced human emotions data
Introduction
Emotion recognition and prediction form a well-researched area in the field of psychology (Cowie et al., 2001). Recently, emotion recognition has drawn growing attention as an evolving research area (Acharya et al., 2020, Chen et al., 2019, Hossain and Muhammad, 2019). Promising results for application areas like online gaming (Kim, Lee, & Provost, 2013), emotion-aware health care systems (Chen, Zhang, Qiu, Guizani, & Hao, 2018), effectiveness based learning programs (Arroyo et al., 2009), digital advertisements, and online shopping (Hossain & Muhammad, 2017) are there. Response to stimulation lasting for seconds or minutes is termed as emotion, the mood is a condition of emotion lasting for hours and days, and a person is inclined to experience certain feelings during that time (Picard, Vyzas, & Healey, 2001). Emotions are an integral element of humans because they motivate action and contribute significance to nearly all human experiences (Sourina, Wang, Liu, & Nguyen, 2011). In the communication between people also, emotions play a crucial part as it affects the contextual comprehension of texts expression or graphics type (Wagner, Kim, & André, 2005).
Based on individual differences in temperament, a dimensional model is proposed to describe their emotional state and attention level (Eerola and Vuoskoski, 2011, Gable and Harmon-Jones, 2010). Another reason is that distinct attention levels can have distinct effects on emotional experiences. A beneficial method for regulating mind and body is attention. The relationships between emotion and attention is a very popular research area for several purposes. One reason is the apparent fact that our emotions can regulate the level of attention (Holland and Gallagher, 1999, Kuusinen et al., 2018, Gordon et al., 2018). While recognizing the emotions of a person, their attention level can also be analyzed. In the traditional way of teaching face to face, instructions were given generally, and by observing students expressions, attentiveness of students can be checked. In the case of online learning; however, these problems become more intense because the instructor finds it hard to distinguish the student’s attention as they are at remote locations and taking courses online. Thus, indirect attention measurement is a highly helpful method that provides teacher feedback in real-time. Learning outcomes are directly affected by the attentiveness of learners (Klimesch et al., 1998, Chen et al., 2019). This study is also addressing the effect of emotions on attention level so that it can be identified that under what circumstances the attention level is enhanced or drops in the participants. As a result, elevated levels of attention offers new ways to learn more and to work more productively also.
In affective computing field, research on recognizing emotions are either based on self-assessment manikin (SAM) that contains subjective measures (De Moor et al., 2006, Bradley and Lang, 1994) or by analyzing physiological and behavioral signals. On the other hand, significant research has been reported using images (Busso et al., 2004, Mollahosseini et al., 2016), speech (Anagnostopoulos, Iliou, & Giannoukos, 2015), audio (Kim & André, 2008), and facial expression (Zheng, 2014). All conventional emotion prediction models for psychological analysis tend to rely on self-assessment forms and human inputs (Kartelj, Filipović, & Milutinović, 2012). To recognize emotions from images, ImageNet was used with the convolutional neural network (CNN) in which static images are considered, and the concept of transfer learning was applied (Ng, Nguyen, Vonikakis, & Winkler, 2015). Kim et al. (2013) used Deep belief network to recognize emotions from speech and video, where he proposed a feature selection technique to enhance the accuracy of the model. Multi-view facial images are also considered for facial expression based emotion recognition by Zhang et al. (2016). Although several previous works are published on audio-visual recognition in the literature, most of them suffer from low accuracy of recognition.
Recently, to recognize emotions using periphery physiological signals has gained much attention in this field (Nakisa et al., 2018, Huang et al., 2019, Zhang et al., 2020). In this category, signals are recorded from brain using electroencephalograph (EEG) and functional magnetic resonance imaging (fMRI), skin temperature (SC), pulse rate, and heart rate that is measured by electrocardiogram (ECG) and electrocorticography (ECoG) (Li and Chen, 2006, Dahl et al., 2014). Among these physiological signals brain, stimuli based signals have gained more popularity. EEG signals have been used widely in medicine to study the brain. EEG is believed to be the finest way of recording data in various modalities owing to its distinctive elements when interacting with emotional conditions (Lokannavar, Lahane, Gangurde, & Chidre, 2015).
It is already proven that humans are adept at concealing and suppressing emotions. Due to which the works published in literature based on the self-assessment test achieved low accuracy in recognizing emotions. Thus, this brings forth the need to develop a new emotion recognition approach. In this study an enhanced fitness function in genetic programming (GP) (Devarriya, Gulati, Mansharamani, Sakalle, & Bhardwaj, 2020) approach to recognize emotions in response to emotional clips is proposed. Emotional movie clips are perceived as a compelling medium for eliciting human emotions, as they are described as the language of emotions (Liang, Oba, & Ishii, 2019). A deep connection among emotion of humans and emotional movie clips is described as the part in the brain of humans that feel and interpret emotions are near together (Pouyanfar & Sameti, 2014).
The challenge in the field of recognizing emotion is the manipulation of emotions. The exact process of recognizing emotions is very time consuming, so there is a need to automate this process, which can help a person to utilize his/her potential to the maximum. It is impossible to falsify brain activity with the help of recording brain signals. The use of portable EEG device is non-invasive, fast, and inexpensive, making it a preferred method to study the response of the brain to emotional stimuli (Acharya et al., 2020) this helps us in achieving a right level of accuracy in the classification of emotions also. EEG signals are recorded using electrodes placed on the scalp of the brain. EEG inputs consist of sample frequency ranges from 0.5 Hz to 100 Hz, and in literature band 0.5 Hz–40 Hz is used for examination of human brain-behavior (Bhatti, Majid, Anwar, & Khan, 2016). EEG signals are classified as five leading frequency bands in this study: Delta, Theta, Alpha, Beta, and Gamma. The other problems in this emotion recognitions dataset is that the dataset is unbalanced. In literature, it is already proven that it is more difficult to classify an unbalanced dataset.
The dataset collected for emotion recognition was unbalanced, as there are more instances of positive emotions than negative ones. An unbalanced dataset is one in which one class have a higher number of samples called Majority class, and other classes have very less number of samples called minority class (Beyan annd Fisher, 2015). In most frameworks, accuracy is used as a measure to assess the results (Acharya, Billimoria, Srivastava, Goel, & Bhardwaj, 2020). However, accuracy provides misleading results in the context of classification of unbalanced dataset (Devarriya et al., 2020). As, a result, the classifier produces higher accuracy for majority class as it influences the classifier because of more number of samples, and lower accuracy for minority classes. A framework that can cope with such situations is therefore required. In the past few years evolutionary algorithms have gained importance in the research field among the researchers for handling classification problems (Chawla et al., 2004, Garcı et al., 2012, Beyan annd Fisher, 2015).
Genetic Programming (GP) (Koza, 1992) inspires the proposed approach to handle classification of unbalanced dataset classification. GP is widely used for different types of classification problems like object detection (Kishore, Patnaik, Mani, & Agrawal, 2000), in medical field (Bhardwaj, Sakalle, Bhardwaj, & Tiwari, 2019), for feature extraction (Bhardwaj et al., 2015, Bhardwaj and Tiwari, 2013, Bhardwaj and Tiwari, 2013), (Guo, Rivero, Dorado, Munteanu, & Pazos, 2011), in game theory (Hargraves and Paris, 1987, Bhowan et al., 2013). GP uses the basic concept of creating multiple functions known as a population and then randomly combine them to create programs. Therefore, in this paper, we introduce an enhanced version of D-score (Devarriya et al., 2020) named as eD-score, to address the issue of unbalanced emotion recognition dataset by using the GP framework. D-score is a fitness function which handles unbalanced dataset being unswayed. Instead of traditional machine learning algorithms which are more lean towards majority classes, D-score gives both the class equal value even though minority classes have less number of samples. The framework in which we implemented eD-score is eDGP.
To show the supremacy of our eDGP we also analyzed it over the two publicly available benchmark dataset DEAP (Koelstra et al., 2011) and SEED (Zheng & Lu, 2015) and compare our work with other state-of-the-art methods as shown in Table 16 (Bhatti et al., 2016, Koelstra et al., 2010, Bastos-Filho et al., 2012). Fast Fourier Transformation is used for feature extraction. We aim to enhance classification results on the state-of-the-art techniques by using an enhanced DGP approach to recognize unbalanced emotions. To illustrate the dominance of our method, we use the 50-50, 60-40, 70-30, 10-fold cross-validation for confusion matrix and Mann Whitney test. Also, evaluated the maximum, minimum, and average classification accuracy of our model. In terms of the social behaviour analysis genre-wise classification of emotions, age group response to emotions, attention level of participants, and gender-based responsiveness is analyzed. The contribution of this paper can be summarized as below:
- •
An enhanced distance score (eD-score) fitness function is proposed to handle unbalanced dataset classification.
- •
A new EEG signal dataset for emotions is created with a portable, single-channel EEG headset (NeuroSky MindWave2).
- •
A novel eDGP framework is proposed for the classification of unbalanced emotion recognition data.
- •
A genre-wise classification, age-group, and gender-based responsiveness of emotions is performed.
- •
Analysis of attention level in participants and its correlation with emotions is done.
The rest of the paper is organized as follows: Section 2 contains related work followed by the background description in Section 3. The proposed methodology is explained in Section 4. Section 5 illustrates the experimental settings followed by performance parameters described in Section 6. Results and discussion is presented in Section 7, and Section 8 contains conclusion.
Section snippets
Related work
A lot of work has been done in recent years for recognizing emotions using EEG signals. Research-based on human emotion recognition based on EEG is mainly divided into three parts, i.e., time domain, frequency domain, and time frequency-domain features (Gupta et al., 2018, Velchev et al., 2016, Ismail et al., 2016). For unbalanced data classification also, some research work is done in the past as most of the real-world data is unbalanced (Duan et al., 2016, Pant et al., 2019). Few of the
Background
This section describes the statutory background for this approach, which includes the feature extraction process fast fourier transformation (FFT) (Zhao & He, 2013) and D-score fitness function (Devarriya et al., 2020). The brief description of FFT and D-score are described next.
Proposed enhanced D-score
This section explains the details of the enhanced D-score fitness function implemented using GP framework as eDGP. We are suggesting an enhanced fitness function, eD-score to solve problems linked to the classification of unbalanced data. Tanh activation function is used in this enhanced fitness function. As, the bounding condition for tanh is [−1,1] so the blowing up issue of activation function is resolved in this. As, DGP makes use of sigmoid function so the problem of blowing up is there.
Experimental setting
In this section, we describe the experimental settings, including stimuli, subjects description, dataset description, and description of the experimental procedure.
Performance measure
The efficiency and reliability of the eDGP classifier area deliberated by performing analytical analysis like confusion matrix, accuracy, F1score, and Mann-Whitney test. Description of these performance measures are as follows:
Results and discussion
This section presents the experimental results of the enhanced D-score Genetic Programming (eDGP) for emotion recognition by the analysis of the EEG signal to validate the effectiveness of our method. Computing environment have Python (3.6) to implement MLP, KNN, SVM, DGP, and eDGP framework as a classifier on a Windows-based operating system having core processor as Intel i7 7th gen of 3.4 GHz with 32 GB of RAM. The feature extraction technique used for all these classifiers is FFT. For the
Conclusion
In the duration of this work, we propose brain signals based human emotion recognition system using enhanced D-score genetic programming, which is analyzing the emotional conditions of participants through brain wave assessment. Our system consists of four different modules: emotion elicitation, EEG data collection, feature extraction, and emotion classification. For this, an eDGP classifier was used for classification after feature extraction through FFT, which is responsible for the resulting
CRediT authorship contribution statement
Divya Acharya: Conceptualization, Investigation, Data curation, Resources, Methodology, Writing - original draft, Software, Formal analysis, Writing - review & editing, Validation, Visualization. Nandana Varshney: Data curation, Formal analysis. Anindiya Vedant: Formal analysis, Software. Yashraj Saxena: Data curation, Formal analysis. Pradeep Tomar: Investigation, Project administration, Supervision, Validation. Shivani Goel: Conceptualization, Methodology, Project administration, Supervision,
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (84)
- et al.
Emotion recognition using fourier transform and genetic programming
Applied Acoustics
(2020) - et al.
Improving bci-based emotion recognition by combining eeg feature selection and kernel classifiers
Expert Systems with Applications
(2016) - et al.
Classifying imbalanced data sets using similarity based hierarchical decomposition
Pattern Recognition
(2015) - et al.
Human emotion recognition and analysis in response to audio music using brain signals
Computers in Human Behavior
(2016) - et al.
Measuring emotion: the self-assessment manikin and the semantic differential
Journal of Behavior Therapy and Experimental Psychiatry
(1994) - et al.
Classification of emotions induced by music videos and correlation with participants’ rating
Expert Systems with Applications
(2014) - et al.
Effects of music tempo and task difficulty on multi-attribute decision-making: An eye-tracking approach
Computers in Human Behavior
(2009) - et al.
Regular exercise, anxiety, depression and personality: A population-based study
Preventive Medicine
(2006) - et al.
Unbalanced breast cancer data classification using novel fitness functions in genetic programming
Expert Systems with Applications
(2020) - et al.
A new support vector data description method for machinery fault diagnosis with unbalanced datasets
Expert Systems with Applications
(2016)
Automatic feature extraction using genetic programming: An application to epileptic eeg classification
Expert Systems with Applications
Amygdala circuitry in attentional and representational processes
Trends in Cognitive Sciences
Emotion recognition using deep learning approach from audio–visual emotional big data
Information Fusion
Induced alpha band power changes in the human eeg and attention
Neuroscience Letters
An unsupervised eeg decoding system for human emotion recognition
Neural Networks
Evolutionary computation algorithms for feature selection of eeg-based emotion recognition using mobile sensors
Expert Systems with Applications
Emotional state classification from eeg data using machine learning approach
Neurocomputing
Emotion recognition using multi-modal data and machine learning techniques: A tutorial and review
Information Fusion
A novel fitness function in genetic programming to handle unbalanced emotion recognition data
Pattern Recognition Letters
Eeg-based automatic emotion recognition: Feature extraction, selection and classification methods
Emotions recognition using eeg signals: A survey
IEEE Transactions on Affective Computing
Features and classifiers for emotion recognition from speech: A survey from 2000 to 2011
Artificial Intelligence Review
A real-time model based support vector machine for emotion recognition through eeg
Evaluation of feature extraction techniques in emotional state recognition
Classification of electroencephalogram signal for the detection of epilepsy using innovative genetic programming
Expert Systems
A novel genetic programming based classifier design using a new constructive crossover operator with a local search technique
Performance improvement in genetic programming using modified crossover and node mutation
An analysis of integration of hill climbing in crossover and mutation operation for eeg signal classification
Evolving diverse ensembles using genetic programming for classification with unbalanced data
IEEE Transactions on Evolutionary Computation
A database of german emotional speech
Ninth European conference on speech communication and technology
Analysis of emotion recognition using facial expressions, speech and multimodal information
Emotion assessment from physiological signals for adaptation of game difficulty
IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans
Special issue on learning from imbalanced data sets
ACM Sigkdd Explorations Newsletter
Emotion recognition from surface emg signal using wavelet transform and neural network
A hierarchical bidirectional gru model with attention for eeg-based emotion classification
IEEE Access
Spha: Smart personal health advisor based on deep analytics
IEEE Communications Magazine
What is the fast fourier transform?
Proceedings of the IEEE
Emotion recognition in human-computer interaction
IEEE Signal Processing Magazine
Mothers’ tone of voice depends on the nature of infants’ transgressions
Emotion
A comparison of the discrete and dimensional models of emotion in music
Psychology of Music
The motivational dimensional model of affect: Implications for breadth of attention, memory, and cognitive categorisation
Cognition and Emotion
Cited by (13)
A new data augmentation convolutional neural network for human emotion recognition based on ECG signals
2022, Biomedical Signal Processing and ControlCitation Excerpt :In our study, we apply both 70–30 technique and 10-fold cross-validation technique. 70–30 technique is performed to illustrate the importance of training data and how vital training information is in the proposed classifier, in this technique, the dataset is split into 70% train data and 30% test data [70]. The 10-fold cross validation technique is used to estimate the skill of a machine learning model on unseen data [71].
A LSTM based deep learning network for recognizing emotions using wireless brainwave driven system
2021, Expert Systems with ApplicationsCitation Excerpt :The physiological signal provides a better understanding of the participant’s reactions generated at the time of experiments. In this category, brain signals are recorded using electroencephalograph (EEG) (Acharya et al., 2021; Lin et al., 2010), functional magnetic resonance imaging (fMRI) (Phan et al., 2002), and electrocorticography (ECoG) (Rao et al., 2018). Among these, EEG signals have gained more popularity in recent years.
EEG based personality prediction using genetic programming
2023, Asian Journal of ControlHuman Behavior Recognition Based on CNN-LSTM Hybrid and Multi-Sensing Feature Information Fusion
2023, Journal of Combinatorial Mathematics and Combinatorial ComputingDeep Learning Based Emotion Recognition Using EEG Signal
2023, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)A Hybrid Model for Classification of Skin Cancer Images after Segmentation
2023, International Journal of Image and Graphics