Multi-label feature selection algorithm based on label pairwise ranking comparison transformation

Multi-label classification refers to the learning problem that a single training sample possibly has multiple labels at the same time. Many real world applications consist of high-dimensional feature vectors, which generally involve some irrelevant and redundant features. This possibly reduces classification performance and increases computational costs. Therefore, feature selection becomes an indispensable pre-processing step. Nowadays filter-type feature selection algorithms based on problem transformation strategies (for example, binary relevance) have attracted more attention due to their high computational efficiency and good classification performance. In this paper, according to the definition of ranking loss, we propose a label pairwise comparison transformation method (PCT), which converts each original multi-label sample into multiple samples with same feature vectors and different label vectors. Further, when PCT is combined with chi-square statistics, we introduce a fast implementation procedure, whose time complexity is approximated to that of binary relevance method. The experimental results of four text data sets show that our proposed algorithm outperforms five existing filter-type feature selection techniques based on problem transformation strategies according to six instance-based evaluation measures.