Graph Quality Judgement: A Large Margin Expedition

Graph as a common structure of machine learning, has played an important role in many learning tasks such as graph-based semi-supervised learning (GSSL). The quality of graph, however, seriously affects the performance of GSSL; moreover, an inappropriate graph may even cause deteriorated performance, that is, GSSL using unlabeled data may be outperformed by direct supervised learning with only labeled data. To this end, it is desired to judge the quality of graph and develop performance-safe GSSL methods. In this paper we propose a large margin separation method LEAD for safe GSSL. Our basic idea is that, if a certain graph owns a high quality, its predictive results on unlabeled data may have a large margin separation. We should exploit the large margin graphs while keeping the small margin graphs (which might be risky) to be rarely exploited. Based on this recognition, we formulate safe GSSL as Semi-Supervised SVM (S3VM) optimization and present an efficient algorithm. Extensive experimental results demonstrate that our proposed method can effectively improve the safeness of GSSL, in addition achieve highly competitive accuracy with many state-of-the-art GSSL methods.