Reconsidering Feature Structure Information and Latent Space Alignment in Partial Multi-label Feature Selection

TL;DR

The paper tackles label disambiguation in partial multi-label learning by introducing latent-space alignment (PML-FSLA). By decomposing the feature and label matrices into latent factors and enforcing alignment between their latent projections, plus a novel $QR$-based feature-selection term, the method leverages accurate feature information to mitigate label-noise and emphasize positive labels. An OPTICS-driven latent dimension $k$ guides the latent space, and a nonnegative, multiplicative-update optimization yields a robust, interpretable solution. Empirical results across eight diverse datasets show PML-FSLA consistently improves macro/micro F1 scores and related metrics, validating its effectiveness in high-dimensional, noise-prone partial labeling scenarios.

Abstract

The purpose of partial multi-label feature selection is to select the most representative feature subset, where the data comes from partial multi-label datasets that have label ambiguity issues. For label disambiguation, previous methods mainly focus on utilizing the information inside the labels and the relationship between the labels and features. However, the information existing in the feature space is rarely considered, especially in partial multi-label scenarios where the noises is considered to be concentrated in the label space while the feature information is correct. This paper proposes a method based on latent space alignment, which uses the information mined in feature space to disambiguate in latent space through the structural consistency between labels and features. In addition, previous methods overestimate the consistency of features and labels in the latent space after convergence. We comprehensively consider the similarity of latent space projections to feature space and label space, and propose new feature selection term. This method also significantly improves the positive label identification ability of the selected features. Comprehensive experiments demonstrate the superiority of the proposed method.

Figures (5)

• Figure 1: An example of partial multi-label learning. The image is partially labeled by noisy annotators. Among the candidate labels, house, street lamp, people and bike are ground-truth labels while tree, mountain and flower are noisy labels.
• Figure 2: The Process of PML-FSLA. First, the feature matrix and the label matrix are projected into the $k$-dimension space determined by OPTICS. Then noisy labels are removed through latent space alignment. Finally two weight matrices are employed for feature selection.
• Figure 3: Through alignment of labels and features in latent space, noisy labels can be found and eliminated due to the structural inconsistencies
• Figure 4: Ten methods on Corel5K in terms of Marco-F1, Average Precision, Ranking Loss and Coverage Error.
• Figure 5: Parameter sensitivity studies on the LLOG_F in terms of Coverage.