FCBF3Rules: A Feature Selection Method for Multi-Label Datasets

In this paper, a novel multi-label feature selection algorithm is introduced based on fast correlation-based filter (FCBF) feature selection method, which is a filter approach for single-label datasets. The strategy of FCBF is that first, it eliminates the features that are irrelevant to classes. Unlike many filter methods which stop on this step, FCBF finds redundant features among relevant features that are remained from the previous step and eliminates them. Therefore, this is one of the most successful single-labels methods in finding the most effective and the smallest feature subset. Extending the step of finding the relevant features in multi-label datasets is not a difficult task. However, in the step of eliminating redundant features, FCBF may removes many effective features. It should be noticed that in multi-label datasets, one feature may be able to distinguish samples that are relevant to a label while another feature is suitable for another label. Hence, these two features cannot be considered to be redundant and one of them cannot be removed. The main contribution of this paper corresponds to the step in which effective and useful features are distinguished from redundant ones in FCBF method. To do so, three rules are implemented and when even one of these rules is not fulfilled, the feature is not removed. The proposed method along with three recently proposed multi-label feature selection methods are applied on 6 standard multi-label datasets for evaluation. The obtained results indicate the strong capability of the proposed algorithm to find the best feature subset, compared to other algorithms.