Infinite Multi-Label Feature Selection

Multi-label feature selection deals with the problem of dimensionality reduction of data in which an instance may belong to multiple class labels simultaneously. Because of computational concerns the existing multi-label feature selection algorithms are not able to consider all possible subsets of feature space in evaluating a candidate feature. This paper proposes a new approach that is able to consider all possible subsets of feature space in evaluating a feature. The proposed method uses the centrality concept in graph theory and reducts the feature evaluation function to finding path costs in feature adjacency graph. Experimental results demonstrate the superiority of the proposed method against state-of-the-art information-theoretical-based filter multi-label feature selection algorithms.

Keywords:

Feature Selection , Multi-Label Learning