Fragmentary label distribution learning via graph regularized maximum entropy criteria

Abstract Label distribution learning (LDL) is a new learning paradigm, which assumes that the labels are related to each instance to some degree. It has been successfully applied in many scenarios due to its ability to tackle label ambiguity. Nevertheless, LDL has also enlarged the labeling costs and difficulties. In many real application areas, such as pattern recognition and image classification, the labeling information is often incomplete, i.e., we cannot determine each label degree to each instance. The performance of traditional LDL algorithm will degrade since they often require complete supervised information. In this paper, we have proposed fragmentary LDL algorithm via Graph Regularized Maximum Entropy criteria (GRME) to solve this problem. It explores the relationship among the labels to recover missing label factors, together with a classifier for categorization. The integration of these two components facilitate the performance enhancement of GRME. Besides, compared with most of traditional methods which are transductive, another advantage of our method is the inductive nature to predict for new coming data directly. We have also extended ADMM algorithm to fit this particular nonconvex problem. The effectiveness of our method is verified by the experimental results on real-world data sets with different settings.