Multi-View Multi-Label Learning With View-Label-Specific Features

Jun Huang, Xiwen Qu, Guorong Li, Feng Qin, Xiao Zheng, Qingming HuangIEEE Access 2019 · OpenReview

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:

  1. Feature Redundancy: Assuming all features in a view are equally relevant to every label.
  2. View Disparity: Treating all data sources as equally informative, despite semantic differences.
  3. 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 VLSF framework for multi-view multi-label learning.

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.

F1 scores of different views on the pascal07 dataset, highlighting view disparity.

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

Abstract

In multi-view multi-label learning, each object is represented by multiple data views, and belongs to multiple class labels simultaneously. Generally, all the data views have a contribution to the multi-label learning task, but their contributions are different. Besides, for each data view, each class label is only associated with a subset data features, and different features have different contributions to each class label. In this paper, we propose a novel framework VLSF for multi-view multi-label learning, i.e., multi-view multi-label learning with View-Label-Specific Features. Specifically, we first learn a low dimensional label-specific data representation for each data view and construct a multi-label classification model based on it by exploiting label correlations and view consensus, and learn the contribution weight of each data view to multi-label learning task for all the class labels jointly. Then, the final prediction can be made by combing the prediction results of all the classifiers and the learned contribution weights. The extensive comparison experiments with the state-of-the-art approaches manifest the effectiveness of the proposed method VLSF.
BibTeX
@inproceedings{huang2019multi,
  author = {Jun Huang and Xiwen Qu and Guorong Li and Feng Qin and Xiao Zheng and Qingming Huang},
  title = {{Multi-View Multi-Label Learning With View-Label-Specific Features}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=wK_zUx8azQ},
}
Figure