Evaluating Feature Selection Methods for Multi-Label Text Classication

Multi-label text classification deals with problems in which each document is associated with a subset of categories. These documents often consist of a large number of words, which can hinder the performance of learning algorithms. Feature selection is a popular task to find representative words and remove unimportant ones, which could speed up learning and even improve learning performance. This work evaluates eight feature selection algorithms in text benchmark datasets. The best algorithms are subsequently compared with random feature selection and classifiers built using all features. Results agree with literature by finding that well-known approaches, such as maximum chi-squared scoring across all labels, are good choices to reduce text dimensionality while reaching competitive multi-label classification performance.