Emotion Recognition Based on Electroencephalogram Using a Multiple Instance Learning Framework

Xiaowei Zhang, Yue Wang, Shengjie Zhao, Jinyong Liu, Jing Pan, Jian Shen, Tingzhen Ding

Abstract: Electroencephalogram (EEG)-based emotion recognition has been widely researched in the field of affective computing. Nevertheless, EEG signals which reflect brain activity are always unstable, it is inappropriate for traditional analysis methods to treat each sliding time window of signals as independent sample during classification. In this study, we employ a multi-instance learning (MIL) framework for EEG-based emotion recognition and regard sliding time windows from the same EEG signal as a whole by learning two MIL models based on Citation-kNN and mi-SVM algorithms. Experiment results show that our methods can achieve higher classification accuracy of 74.21% and 77.50% on two affective dimensions (valence and arousal) respectively when comparing with traditional single-instance classification algorithms. We believe that MIL framework can improve the generalization performance of EEG-based emotion recognition further, and provide new inspiration for affective computing.