Feature selection techniques with class separability for multivariate time series

Feature selection is very important in the mining of multivariate time series data, which is represented in matrix. We propose a novel filter method termed as class separability feature selection (CSFS) for feature selection from multivariate time series with the trace-based class separability criterion. The mutual information matrix between variables is used as the features for classification. And the feature selection algorithm CSFS selects features according to the scores of class separability and variable separability. The proposed method is compared with CLeVer, Corona and AGV on the UCI EEG data sets, and the simulation results substantiate the good performance of CSFS.