PHFS: Progressive Hierarchical Feature Selection Based on Adaptive Sample Weighting

Abstract

Hierarchical feature selection is considered an effective technique to reduce the dimensionality of data with complex hierarchical label structures. Incorrect labels are a common and challenging issue in complex hierarchical data. However, the existing hierarchical methods often struggle to dynamically adapt to label noise and lack the flexibility to adjust sample weights. Therefore, their effectiveness in managing complex data with many classes and mitigating label noise is significantly limited. To address these issues, in this article, an adaptive sample weighting-based progressive hierarchical feature selection (PHFS) method was proposed, which dynamically adjusts the sample weights to focus on high-quality data. PHFS integrates progressive sample selection and hierarchical feature selection into a unified framework, thus enhancing its effectiveness in reducing the impact of label noise and achieving optimal performance. The progressive selection process is divided into initial and subsequent stages, focusing on correct and incorrect samples. In the initial stage, PHFS selects valuable and correct samples based on the adaptive weights calculated through hierarchical classification feedback, maximizing the guiding effect of the correctly labeled examples. In the subsequent stages, PHFS uses matrix factorization to preserve the structure of the correctly labeled samples, preventing the forgetting of the early selected samples and minimizing the negative impact of the mislabelled samples. The superiority of PHFS over 13 state-of-the-art methods was demonstrated by performing extensive experiments on eight real-world datasets, highlighting its effectiveness in reducing label noise and achieving optimal performance.