Multi-label Learning Based On Label-specific Feature Extraction

In the framework of multi-label learning, each instance is represented by a feature vector and is simultaneously assigned with more than one class label. Multi-label data usually present the characteristics of high dimension, much redundant information, and so on, which make dimensionality reduction technology more and more important in multi label learning. Since different class labels may have their own unique characteristics, they are called label-specific features. Based on the above assumption, we propose a multi-label learning approach with label-specific features called MLLSFE to extract low dimensional features for all labels. The proposed algorithm implements the label-specific feature extraction by the thought of pairwise constraint dimensionality reduction. Extensive experimental results conducted on different datasets show that the proposed algorithm can effectively promote the classification performance in multi-label learning.