GraphDPI: Partial label disambiguation by graph representation learning via mutual information maximization

Abstract

Partial label learning (PLL) is a weakly supervised learning framework where each training instance is associated with more than one candidate label, and only one of them is the true label. Most of the existing PLL algorithms directly disambiguate the candidate labels according to the instance feature similarity, but fail to discover the latent semantic relationship over the entire dataset. In this paper, method GraphDPI, an innovative deep partial label disambiguation by graph representation via mutual information maximization, is proposed. This method can capture the semantic clusters with the most unique information in the latent space and automatically adapt to different feature distributions. Specifically, a new sampling method based on the graph is proposed to estimate mutual information, extending GCN to the field of weakly supervised learning. Therefore, the graph representation of the data can contain more distinguishing information to disambiguate candidate labels by maximizing the mutual information of the local graph representation and the global one. Furthermore, the triplet loss is introduced to fully exploit the relationship between instances and extract the latent embedding representation over the entire dataset. It thereby can make the model output as large as possible on the inter-class variation and as small as possible on the intra-class variation. Finally, the candidate labels can be disambiguated by the difference between semantic clusters. Experiments reveal the overwhelming performances of GraphDPI.