Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning

Pingting Hao, Kunpeng Liu, Wanfu GaoAAAI 2024 · DOI

Introduction

Multi-view multi-label (MVML) learning has become a cornerstone of modern data analysis, where objects are characterized by multiple feature sets (views) and associated with multiple semantic labels. For instance, a multimedia clip can be described by visual frames, audio tracks, and textual metadata. Feature selection in this context aims to identify the most discriminative attributes across all views to mitigate the "curse of dimensionality" and enhance model performance.

However, existing MVML feature selection methods often struggle with two critical issues: label noise and view-specific label relevance. Human-provided labels are frequently noisy, and not all labels are equally relevant to every view. For example, in a music video, the label "guitar" is highly relevant to the audio view but may be irrelevant to a text view discussing the artist's biography.

Illustration of hybrid labels: common (green), noisy (red), and view-specific (black) labels in a multi-view context.

Methodology: The DHLI Framework

To address these challenges, we propose the Double-Layer Hybrid-Label Identification (DHLI) framework. Unlike traditional methods that focus solely on the commonalities between views, DHLI decomposes the observed label matrix into three distinct components:

  1. Common Labels: Information shared across all views.
  2. View-Specific Labels: Information unique to individual views, further refined through a double-layer splitting process.
  3. Noisy Labels: Erroneous labels that are explicitly filtered out to prevent them from degrading the feature selection process.

Technical Innovations

DHLI utilizes a unified loss function incorporating $l_{2,1}$-norm regularization for group sparsity, ensuring that only the most relevant features are selected. A key innovation is the use of logic-based regularization (implemented via Hadamard products) to enforce constraints on the relationship between common, specific, and noisy labels. This ensures that the specific labels of different views do not redundantly overlap and that noise is effectively isolated. The resulting non-convex objective function is solved using derived multiplicative update rules with proven convergence.

Experimental Evaluation

The DHLI framework was evaluated on six benchmark datasets: SCENE, OBJECT, MIRFlickr, Corel5K, IAPRTC12, and 3Sources. These datasets span various domains and scales, providing a comprehensive test of the model's robustness. Performance was measured using four standard metrics: Average Precision, Coverage, Hamming Loss, and Ranking Loss.

Results and Discussion

Experimental results demonstrate that DHLI consistently outperforms seven state-of-the-art methods, including MSFS, ELSMML, and M2LD. As shown in the performance graphs, DHLI (represented by the black line with circular markers) achieves superior results across most metrics and datasets.

Performance comparison of DHLI against state-of-the-art methods across four evaluation metrics.

Ablation studies confirmed that each component of the hybrid-label identification—common, specific, and noisy—is essential for optimal performance. Furthermore, parameter sensitivity analysis indicated that DHLI is stable across a wide range of hyperparameter settings, making it a reliable choice for practical applications.

Conclusion

The DHLI framework represents a significant advancement in MVML feature selection by explicitly modeling the complex "hybrid" nature of labels. By distinguishing between common, view-specific, and noisy information, DHLI achieves a more accurate mapping between features and labels. Future work will focus on scaling the framework for extreme multi-label problems and extending it to handle non-aligned or missing views.

Summary updated

BibTeX
@inproceedings{hao2024double,
  author = {Pingting Hao and Kunpeng Liu and Wanfu Gao},
  title = {{Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i11.29120},
}
Figure