Feature selection Model and Generalization Performance of Two-class Emotion Recognition Systems Based on Physiological Signals
基于生理信号的二分类情感识别系统特征选择模型和泛化性能分析

The feature selection process in an emotion recognition system is an NP hard problem, i. e. the scale of the problem increases exponentially with the increasing number of initial features. I}he goal of establishing good two-class emotion recognition systems was to find a subset of the initial features which minimized the missing rate and the false rate of the system. Such a task was regarded as a combinatorial optimization problem and solved by Tabu search algorithm and the Fisher classifier. Two kinds of physiological signals(the Galvanic Skin Response and the Heart Rate) recorded under four discrete emotion states(joy,anger,grief and fear) of 66 college students were used during the establishment of the systems. It was found that the problem of feature selection could be properly solved by Tabu search, and the user-independent emotion recognition systems had good generalization performance. Furthermore, the individual difference of affective physiological responses had different influence on the recognition of joy,anger,grief and fear.