Reconhecimento de emoções baseado em Aprendizado Autossupervisionado

Abstract

Emotion recognition is a machine learning application that involves analyzing physiological, audio, and/or video signals to identify emotions expressed by individuals. Obtaining labeled datasets for this task is challenging and costly, often presenting structural problems such as class imbalance, missing data, and labeling biases. A promising approach to circumvent these problems is to develop pattern recognition solutions based on self-supervised learning. This approach allows training models using unlabeled data, transferring the acquired knowledge to a model specialized in emotion recognition. In this way, it is possible to overcome the dependency on labeled datasets, making the process more efficient and less costly. The choice of auxiliary tasks in self-supervised learning is crucial, as it enables efficient training of models on large unlabeled datasets and contributes to learning robust and generalizable representations. This allows the model to better adapt to different tasks and scenarios. In this context, this work presents a neural network architecture that uses a self supervised approach for emotion recognition from electrocardiogram signals. To evaluate the performance of the proposed neural architecture, we implemented and evaluated different combinations of auxiliary tasks, analyzing how each one contributes to the model's effectiveness and accuracy. We identified the most significant auxiliary tasks for emotion classification and conducted detailed analyses of the parameters associated with these tasks. Experiments conducted on four public datasets consistently demonstrated the superior performance of the proposed method compared to the same architecture trained in a supervised manner. In the SWELL dataset, the method achieved an accuracy of 93.64% in arousal classification, which is the degree of alertness of the emotion,, using only 25% of the labeled data, compared to 78.20% for the supervised method.