Partial multi-label learning via K-means graph transformer

Partial Multi-Label Learning (PML) is a weakly supervised learning problem in which each instance is assigned a set of candidate labels, including relevant and irrelevant ones. The traditional label disambiguation strategies often rely on prior knowledge or auxiliary information and do not fully consider the complementary information between different label features. To tackle this challenge, we propose a K-means Graph Transformer for PML (PML-KGT), which introduces cluster centers and graph structure design to approximate label features. The goal of this process is to learn the features of different label classes through the complementary information from candidate labels of similar instances in the graph structure, while effective label features can accurately measure the relationship between instance and candidate labels, thus mitigating the influence of noisy labels. Additionally, we introduce a novel partial multi-label correction loss that determines candidate label weights based on evaluating the correlation between cluster centers and instances, thereby alleviating the effect of noise labels. As training progresses, the ground-truth labels are gradually recognized, and the improved cluster centers and labels contribute to enhancing the performance of the classifier. Comprehensive experiments on the real-world and synthetic PML datasets validate the advantage of the PML-KGT.