Feature Selection Based on Streaming Labels Considering the Division of the Selected Feature Set

Feature selection based on streaming labels has received extensive attention in recent years, which is a novel and challenging task for multi-label learning. It assumes that all labels are not given in advance and flow into the system gradually one by one over time. Existing information-theoretical-based feature selection methods with streaming labels use mutual information or conditional mutual information to measure the candidate feature correlation when a new label arrives. However, they do not consider the different influence of already-selected features for candidate feature correlation when facing different new labels. Furthermore, all already-selected features are calculated as having a redundant relationship with candidate features. In fact, some already-selected features are supplementary to candidate features with respect to different new labels. To address this issue, we first redivide the selected features to consider the different effects on the candidate feature correlation and redundancy for the newly arrived label. Second, a feature fusion strategy is used to fuse the subsets of features corresponding to all arrived labels. Finally, the proposed method compares to six other state-of-the-art feature selection methods on ten data sets and employs three evaluation metrics to verify the classification performance. Experimental results show that the proposed method is superior to other methods.