A multi-label learning method for efficient affective detection

Biosignals-based affective computing plays an important role in human-computing interaction. In recent years, a lot of works have been successfully implemented for emotion recognition with biological signals. However most of them are computationally expensive due to the complexity of models. In this paper, we present a multi-label learning (MLL) method to map biological signals to an affective model in real time. Multi-label learning combines multiple classifiers in a same training process by evaluating correlations between labels. To evaluate the proposed model, 25 male subjects were recruited to anticipate an experiment for model evaluation. Dynamic images were specifically chosen to elicit emotion described by different arousal and valence levels in the experiment. In terms of physiological signals, electrocardiogram (ECG) and skin conductivity (SC) signals were collected for classification. By applying the MLL method to analyze the relationships between physiological signal features and affective labels, we observed that the proposed method performed better than traditional method, which showing the potential of MLL in affective detection.