Hybrid feature-based collaborative disambiguation for multi-instance partial-label learning.

Multi-instance partial-label learning (MIPL) deals with the task of learning from multi-instance bags associated with candidate label sets. Each bag contains a set of instances, and only one true label lies within its candidate label set. Existing MIPL methods, such as DEMIPL and MIPLGP, primarily rely on bag-level aggregated features for training and disambiguation. However, the aggregation operation inevitably discards critical instance-level details and reduces feature information density, which limits disambiguation performance. To tackle this issue, we propose a novel MIPL method named HFC-MIPL (Hybrid Feature-based Collaborative Disambiguation for MIPL). We first build multi-granular hybrid features by fusing purified instance representations and aligned bag embeddings to retain both local and global semantics. A dual disambiguation scheme combines focusing regularization for distribution shaping with manifold propagation for pseudo-label refinement through feature similarity. In addition, an epoch-adaptive filtering strategy based on raw sigmoid attention determines the purified instance summary used for graph construction. Experiments on image, audio, and medical datasets show that HFC-MIPL achieves competitive performance against the compared MIPL and PLL baselines across different candidate-label ambiguity settings. Code: https://github.com/hzzhuzhen/HFC_MIPL.