Semi-Supervised Learning by Mixed Label Propagation

Recent studies have shown that graph-based approaches are effective for semi-supervised learning. The key idea behind many graph-based approaches is to enforce the consistency between the class assignment of unlabeled examples and the pairwise similarity between examples. One major limitation with most graph-based approaches is that they are unable to explore dissimilarity or negative similarity. This is because the dissimilar relation is not transitive, and therefore is difficult to be propagated. Furthermore, negative similarity could result in unbounded energy functions, which makes most graphbased algorithms unapplicable. In this paper, we propose a new graph-based approach, termed as “mixed label propagation” which is able to effectively explore both similarity and dissimilarity simultaneously. In particular, the new framework determines the assignment of class labels by (1) minimizing the energy function associated with positive similarity, and (2) maximizing the energy function associated with negative similarity. Our empirical study with collaborative filtering shows promising performance of the proposed approach.