Multi-label learning with kernel local label information

Abstract

It is important to fully utilize label correlations in multi-label learning. If there is a strong positive correlation between label i and label j, an instance associated with label i also likely has label j simultaneously. So, label correlations can provide some auxiliary information when predicting unseen instances. Existing some multi-label algorithms utilize label correlations to constraint model parameters in the training stage, while label correlations are ignored in the prediction stage. Moreover, it is difficult to obtain relatively accurate label correlations by directly observing data when some labels have few positive instances in the training data. In this paper, instead of directly calculating label correlations by cosine distance and so on, we introduce a kernel function and an manifold regularization to learn them by iteratively updating. Meanwhile, we utilize them and local label information to aid label prediction. Ultimately, unseen instances are predicted by combining auxiliary label predictions and the model outputs. We compare the proposed algorithm with related algorithms on 10 data sets, and the experimental results validate its effectiveness.