Partial Label Learning Tailored Graph Construction
Fuchao Yang, Yongqiang Dong, Yuheng Jia · 2025
In partial label learning (PLL), each sample is annotated with a group of candidate labels, among which only one label is correct. The key of PLL is to find the ground-truth label concealed in the candidate label set, which is known as label disambiguation. The instance relationships captured by a graph play a central role in label disambiguation, as if two samples are close to each other in the feature space, they are expected to share the ground-truth same label. However, the existing PLL methods simply use the feature matrix to construct the graph without considering the characteristics of PLL. In this paper, we propose a novel graph construction model that is tailored to PLL. Specifically, we first build a local similarity matrix by reconstructing a sample by its neighbors. Second, we design a dissimilarity matrix to specify the highly dissimilar samples according to the available partial labels, and further enhance it by dissimilarity propagation. As the similarity and the dissimilarity matrices form an adversarial relationship, the enhanced dissimilarity matrix is used to refine the similarity matrix. Then, the proposed model is finally formulated as a dissimilarity propagation guided graph learning problem, which is solved by the inexact augmented Lagrange multiplier method. Extensive experiments on artificial as well as real-world partial label data sets demonstrate that the learned graph can correctly capture the similarity relationships among samples and improve the classification performance of different graph-based PLL methods. The code implementation is publicly available at https://github.com/Yangfc-ML/PL-TGC.