A Hybrid Method to Find Effective Subset of Features in Multi-label Datasets

In multi-label data, each instance is associated with a set of labels, instead of one label. Due to the increasing number of modern applications associated with multi-label data, multi-label classification has gained significant attention during recent years. As in single-label data, eliminating redundant and/or irrelevant features plays an important role in improving classification performance. In this paper, a hybrid method for multi-label feature selection problem based on combing filter and wrapper methods is proposed, where meta-heuristic algorithms are employed as the wrapper method. Since, the number of features in multi-label data is usually high, solely employing search algorithms for finding the optimal feature subsets has high computational burden, and is very possible to fail. Hence, irrelevant features are first detected and removed by a filter method. Then, salient features are found among the remained features by the help of meta-heuristic algorithms. A significant number of well-known meta-heuristic algorithms are employed as the wrapper method, in the proposed system. Experiments show that the proposed method obtains better classification results, compared to other algorithms.