Weakly-supervised label distribution feature selection via label-specific features and label correlation

Abstract

Label Distribution Learning (LDL) stands out as an innovative method to address the challenges posed by label ambiguity. Current LDL algorithms are predominantly designed for datasets with comprehensive supervised information. However, in real-world scenarios, it’s common to encounter partial missing labels within the label space. This phenomenon disrupts the structure and correlation between labels, posing a challenge to the precise design of learning algorithms. Furthermore, in label distribution learning, it also faces the effect of high feature dimensionality. To tackle this, adopting pre-processing methods like feature selection becomes crucial, aiming to trim down the data dimensionality. Motivated by this, a weakly-supervised label distribution feature selection algorithm based on label correlation is proposed in this paper. First, to handle weakly-supervised label distribution data, a two-stage incomplete label distribution learning method (IncomLDLTS) to recover missing labels by exploiting label-independent prediction based on label-specific features is proposed. Second, a minimum correlation label feature selection algorithm (MCLFS) to enhance the performance for complete label distribution data is designed, which employs the interaction information metric to explore the label correlation to obtain a minimum correlation label, then features about each label based on label-specific features and feature redundancy are captured. Finally, based on six representative evaluation metrics, our experiments across 14 datasets affirm the effectiveness of our approach, not only in restoring missing labels but also in choosing essential features, leading to enhanced classification accuracy.