Partial multi-label feature selection based on label distribution learning

Abstract

Partial Multi-label Learning (PML) induces a multi-classifier in an imprecise supervised environment, where the candidate labels associated with each training sample are partially valid. The high-dimensional feature space, presented in PML data accompanied by ambiguous labeling information, is a significant challenge for learning. In this paper, we propose a PML feature selection method based on Label Distribution Learning (LDL), which handles the above challenges by correcting misleading and then selecting common and label-specific features. In the first procedure, the error distribution hypothesis is constructed, which divides the structure of ambiguous label information into minority and majority error distribution according to the error amount that may appear in the data annotation process. Under the analysis of the hypothesis, the label credibility distribution data (LCDD) was generated by identifying and correcting errors, where the fractional category of each label associated with each training sample describes the probability that the label belongs to that sample. In the second procedure, a discriminative feature subset is selected for PML based on LCDD by common and label-specific feature constraints. Experiments on three synthetic and five real PML datasets demonstrate the effectiveness of the proposed method.