A Novel Unsupervised Feature Selection based on Correlation Measure between Heterogeneous Features

Most existing feature selection algorithms only focus on a certain type of data, either numerical or categorical. In this paper, we present an unsupervised heterogeneous feature selection method based on a novel feature correlation measurement. Firstly, a unified correlation measuring framework is designed to quantify the correlations between any two features, including features with the same type and different types. Then, a filter-based feature selection algorithm is proposed to remove redundant features and obtain a best feature subset. Various experiments have been conducted on different mixed data sets and the results indicate that the proposed method can obtain better clustering performance with a smaller heterogeneous feature subset.