.png){kind=icon}
.png){kind=icon}
Return
Partial multi-label learning via adaptive bipartite graph embedding
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.134489.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Partial multi-label learning in which each data example is associated with a set of candidate label sets, of which only a subset are valid ground-truth labels for the training examples. Many graph-based methods have been developed and have achieved satisfactory performance, helping to disambiguate the candidate label sets by learning the graph structure of the instances. Nonetheless, existing graph-based methods have two drawbacks: 1) some models perform similarity graph construction and classification in separate steps and the roughness of the two stages results in the constructed similarity graphs not being optimal for the classification task; 2) most of the models suffer from high computational complexity in constructing pairwise similarity graphs, which is not beneficial for practical applications. This paper proposes a novel PML method, jointly adaptive bipartite graph embedding and label correlation learning. Specifically, anchor samples are mined in the kernel space to construct the similarity graph, and the l2,p -norm is used to mitigate the “anchor shift” problem caused by sample noise. Then, the soft label matrix is introduced to promote its participation in the graph construction process to ensure that the model can restore the confidence of candidate labels at the feature and semantic levels to produce relatively accurate pseudo-labels. Finally, label correlation and bipartite graph feature-induced disambiguation models optimize the classifiers within a unified framework. Extensive experimental studies on real-world and synthetic data validate the effectiveness and robustness of the proposed approach in solving the PML problem.
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W