Label matrix normalization for semisupervised learning from imbalanced Data

Manually labeled data-sets are vital to graph-based semisupervised learning. However, in the real world, labeled data-sets are often heavily imbalanced, and the classifiers trained on such skewed data tend to show poor performance for low-frequency classes. In this paper, we deal with an imbalanced data case of semisupervised learning and propose a novel label matrix normalization solution called LMN to tackle the general imbalance problem. Experiments over different data-sets reveal the effectiveness of the devised algorithm.