Large Margin Based Multi-label Feature Selection

Since multi-label learning studies the problem where each instance is associated with several labels simultaneously, it can be applied in more domains comparing with single label learning and has attracted increasing attention from researchers. However, multi-label learning suffers from the curse of dimension either and hence corresponding feature selection algorithms are required.Feature selection has been deeply studied in traditional single label problem. However, when it comes to multi-label problem, most feature selection algorithms for single label problem are infeasible: 1. Their criterions to evaluate features can not be applied to several labels simultaneously 2. New methods to model label correlations are needed to help improve performance.In this paper, firstly we summarize a framework according to classic multi-label learning algorithms. Then a new measurement is proposed to measure similarity between multi-label instances and hence spectral feature selection framework(SPEC) can be adapted to handling multi-label problem. Finally, we proposed a new large margin based multi label feature selection algorithm(ML_LMBA) for multi-label learning problem. According to similarity between instances, SPEC and ML_LMBA extract information from both feature space and label space, and therefore they may utilize label correlations while they are independent to problem transformation strategy and classifiers. Experiments on real world data sets demonstrate the correctness and high performance of the proposed algorithm.