Classificação de sinais fisiológicos para inferência do estado emocional de usuários

Abstract

Emotional state plays a fundamental role in explaining human behavior in everyday life, influencing decisions and even how to communicate. Therefore, understanding these characteristics and how to identify them is of the utmost importance for a better understanding of human behavior. Emotions can be classified into two models of taxonomy: Discrete Model that include basic emotions (happiness, sadness, fear, disgust, anger, surprise) and Dimensional Model that expresses two or more emotions in a space, with emotional domains such as Valencia (disgust, pleasure) and Excitement (calm, excitement). The literature presents ways of evaluating emotional cues, with inference in real time, by collecting physiological signals through sensors and classified by algorithms in the emotional domains of Valencia and Excitation. In this context, there is a complexity in collecting this data with low-cost sensors, as well as classifying emotions into more emotional domains, increasing the number of classes and consequently the accuracy of the classification in space. This dissertation aimed to classify physiological signals collected with low cost commercial sensors, inferring dimensional emotions in four domains: Valencia, Excitation, Feeling of control and Ease of conclusion of the objective. Thus, this work presents a dataset, with data from three sensors: cardiac activity (ECG), brain activity (EEG) and galvanic response (GSR). To compose the dataset, the physiological signals were collected from 33 participants in three sessions. In order to bring the individual to a desired emotional state, 16 prelabeled movie clips and video clips were used. After the collection and recording of the signals, a preprocessing step was performed to eliminate noise and inconsistent data and extraction of characteristics. In order to classify, we used the closest K-neighbors algorithms, Naive Bayes, Decision Tree, Random Forest, XGBoost, Support Vector Machine and Artificial Neural Networks. Finally, a statistical evaluation of the performance of the algorithms in each sensor data was performed. The classification algorithms that best fit the characteristics of the data were Naive Bayes for ECG and GSR with 96% and 77% accuracy and Support Vector Machine for EEG with 99% accuracy. The results suggest that the data collected allow classification in the four domains studied.