Fuzzy neighborhood-based partial label feature selection via label iterative disambiguation

Abstract

Partial label learning is a specific weakly supervised learning framework in which each training sample is associated with a candidate label set in which the ground-truth label is concealed. Feature selection can remove redundant and irrelevant features to improve the generalization performance of the classification model. However, the impact of ambiguous labels is an essential challenge when adopting feature selection for partial label data. In this paper, a novel two-stage feature selection method is proposed, called fuzzy neighborhood-based partial label feature selection with label iterative disambiguation. In the first stage, the proposed method addresses the issue of noise labels by employing a neighborhood-based iterative strategy to enlarge the gap between ground-truth labels and noisy labels. Subsequently, the labeling confidence induced by label disambiguation is utilized to enhance the robustness of feature selection. In the second stage, feature significance is evaluated using three metrics based on fuzzy neighborhoods. Specifically, fuzzy dependency is obtained using fuzzy relations and labeling confidence. Fuzzy neighborhood entropy-based information gain is proposed as an uncertainty measure. Furthermore, the similarity between samples in the same fuzzy neighborhood is used to estimate neighborhood consistency. The fusion of the above metrics can select more discriminative features for partial label learning. Finally, experimental results on eight controlled UCI datasets and five real-world datasets demonstrate that the proposed method achieves superior performance than other compared methods.