Abstract
In multi-view multi-label learning, each object is represented by multiple data views and belongs to multiple class labels simultaneously. Generally, all data views contribute to the multi-label learning task, but their contributions differ. Furthermore, for each data view, each class label is associated with only a subset of data features. In this paper, we propose a novel framework, VLSF (View-Label-Specific Features), for multi-view multi-label learning. VLSF learns low-dimensional label-specific data representations for each view, exploits label correlations and view consensus, and jointly learns the contribution weight of each data view. Extensive experiments demonstrate that VLSF outperforms state-of-the-art approaches.
Introduction
Real-world data is increasingly complex. A single object, such as a news article or a digital image, is often described by multiple "views" (e.g., color histograms, textures, and user-generated tags) and associated with multiple "labels" (e.g., "beach," "sunset," and "vacation"). Traditional multi-view multi-label learning (MVML) methods often struggle with three key issues:
- Feature Redundancy: Assuming all features in a view are equally relevant to every label.
- View Disparity: Treating all data sources as equally informative, despite semantic differences.
- Ignoring Relationships: Failing to model correlations between labels and consensus between views.
The VLSF Framework
To address these gaps, the researchers developed the VLSF framework. This approach integrates three critical components into a unified learning model:
- Label-Specific Representations: By employing $\ell_1$-norm regularization, VLSF identifies a sparse subset of features within each view that are most discriminative for specific labels. This reduces noise and prevents overfitting by focusing on relevant data.
- Consensus and Correlation: The model utilizes a graph Laplacian based on cosine similarity to capture label correlations. Simultaneously, it enforces a flexible view consensus, assuming that views with similar importance should yield consistent predictions.
- Adaptive View Weighting: VLSF automatically learns a weight ($\theta_v$) for each data view, allowing it to prioritize highly informative sources (like semantic tags) over noisier ones.
The resulting objective function is optimized using the Accelerated Proximal Gradient (APG) method, ensuring efficient convergence for large-scale tasks.
Experimental Results and Insights
The authors validated VLSF on 12 dataset configurations, including benchmarks like pascal07, corel5k, and mirflickr. VLSF consistently achieved superior accuracy across multiple metrics, including Hamming Loss and Average Precision, compared to state-of-the-art methods like LLSF and iMVML.
The Impact of View Disparity
A key insight from the study is the significant disparity in the predictive power of different data views. For instance, on the pascal07 dataset, high-level semantic "Tags" are far more informative than low-level visual features like "Gist" or "RGB" color spaces.
By adaptively weighting these views, VLSF effectively "listens" more to the most reliable sources of information. Ablation studies further confirmed that the combination of view weighting, label correlation, and view consensus is essential for the model's high performance.
Conclusion
VLSF provides a robust solution for MVML by intelligently identifying label-specific features and weighting the importance of diverse data views. While currently based on linear relationships, future work may explore non-linear kernels or deep learning architectures to further enhance the framework's ability to capture complex data patterns.
Summary updated