Feature Selection Method Based on FSA-Choquet Fuzzy Integral

Abstract

With the instant development and wide application of sensor technology, human behavior recognition based on wearable sensors has attracted wide attention and become a hot research topic. The typical Choquet fuzzy integral feature selection method considers the interaction between classes, but the preferred features are not tested in the feature selection process. The selected feature subset is not suitable for different classifiers and different classification actions, and there are redundant features. Consequently, this paper proposes an adaptive feature selection method named Feature Selection based on the Adaptive Choquet fuzzy integral (FSA-Choquet). This proposed method can get the optimal subset by two times of feature selections. First, Choquet fuzzy integral is applied to optimize the current optimal feature subset. Then the feature selection is processed by combining the maximum redundancy calculation with the backward floating search strategy and classifier. At last, Experiments show that the feature based on FSA-Choquet selection has a higher classification recognition rate in behavior recognition. Under the comparison based on SVM classifier, the recognition accuracy of the features selected by FSA-Choquet is as high as 89.6 %, and the recognition rate of the features selected by Choquet is 77.36 %. Besides, Under the classifier based on Naive Bayes (NB), the classification recognition rate of FSA-Choquet is 91.67 %, while that of Choquet is only 82.3 %.