Incomplete Multi-view Partial Label Feature Selection

Partial-label feature selection can simultaneously disambiguate candidate labels and select discriminative features when each training instance is associated with multiple candidate labels, and it has proved effective for conventional single-view partial-label problems. However, data in many real-world applications are naturally described by multiple views and may contain missing views, rendering existing methods inadequate. To address this issue, we propose IMPLFS, an incomplete multi-view partial-label feature selection method based on complementary learning.Three key issues need to be addressed: how to learn reliable sample similarities from incomplete views, how to fully exploit complementary information across views, and how to reduce the adverse effect of candidate-label ambiguity on feature evaluation. First, IMPLFS develops a complementary-learning-guided similarity-matrix model that uses observed within-view relations and cross-view complementary relations to complete and enhance the sample-similarity structure of each view. Second, it dynamically integrates the complementary information supplied by different views into a fused graph through graph learning and adaptive view weighting. Finally, it constructs a joint optimization framework for label-confidence estimation and feature selection, allowing label disambiguation and feature evaluation to be updated collaboratively. Experimental results on multiple benchmark datasets demonstrate that IMPLFS outperforms competing feature selection methods and can effectively address the incomplete multi-view partial-label feature selection problem.