Multi-label Feature Selection Method Combining Unbiased Hilbert-Schmidt Independence Criterion with Controlled Genetic Algorithm

In multi-label learning, some redundant and irrelevant features increase computational cost and even degrade classification performance, which are widely dealt with via feature selection procedure. Unbiased Hilbert-Schmidt independence criterion (HSIC) is a kernel-based dependence measure between feature and label data, which has been combined with greedy search techniques (e.g., sequential forward selection) to search for a locally optimal feature subset. Alternatively, it is possible to achieve a globally optimal solution using genetic algorithm (GA), but usually the final solution prefers to select about a half of original features. In this paper, we propose a new GA variant to control the number of selected features (simply CGA). Then CGA is integrated with HSIC to formulate a novel multi-label feature selection technique (CGAHSIC) for a given size of feature subset. The effectiveness of our proposed CGAHSIC is validated through comparing with four existing algorithms, on four benchmark data sets, according to four indicative multi-label classification evaluation metrics (Hamming loss, accuracy, F1 and subset accuracy).