A Multi-label feature selection algorithm based on multi-objective optimization

Multi-label performance evaluation metrics could be mainly grouped into two parts: ranking-based and instance-based. The former is based on discriminant function values (e.g., average precision and ranking loss). The latter is associated with predicted relevant label subsets (e.g., Hamming loss and accuracy), which is determined via a proper threshold from the discriminant function values. Firstly, we show that such two parts conflict with each other possibly according to the theoretical and experimental analysis in this study. Therefore a multi-label wrapper feature selection method essentially needs to optimize multiple objective functions. In this paper, given multilabel k-nearest neighbour method, we utilize evolutionary multiobjective optimization algorithm (NSGA-II) to maximize average precision metric and minimize Hamming loss one simultaneously, to construct a novel feature selection approach for multilabel classification. Experiments illustrate that our method could achieve a better performance than the other existing techniques.