Abstract
Label enhancement (LE) is a critical process for recovering nuanced label distributions from binary logical labels in multi-label data. While existing LE methods focus on single-view data, real-world applications often provide multi-view data with rich, heterogeneous semantic information. This paper introduces TMV-LE, a tensor-based framework that leverages multi-view information through tensor factorization to capture high-order relationships. By mining the topological structure of the feature space and migrating it to the label space, TMV-LE significantly improves performance in multi-view multi-label learning tasks.
Introduction
In traditional Multi-Label Learning (MLL), objects are tagged with binary labels (e.g., "yes" or "no"). However, this approach fails to capture the relative importance or intensity of each label—a problem addressed by Label Distribution Learning (LDL). Because LDL datasets are expensive to create, Label Enhancement (LE) was developed to recover these distributions from existing binary data.
Most LE research has focused on single-view data. However, many datasets are inherently multi-view (e.g., an image described by color, texture, and shape). TMV-LE addresses this by integrating these diverse perspectives to create a more accurate and robust label distribution.
The TMV-LE Framework
TMV-LE operates through a two-stage process designed to exploit the consistency and complementarity of multiple data views:
- Common Representation Learning: To handle views with different dimensions, the model expands them into a uniform size and stacks them into a third-order tensor. It then employs CP (CANDECOMP/PARAFAC) factorization to derive a common subspace ($P$). This approach captures complex, non-linear relationships across views that simple matrix concatenation would miss. $l_{2,1}$-regularization is applied to maintain view-specific nuances while extracting shared semantics.
- Topological Mining and Migration: The framework identifies the "shape" or topology of the feature space using Local Linear Embedding (LLE). It operates on the principle that instances close in the feature space should share similar label distributions. This discovered structure is then "migrated" to reconstruct the final label distributions. Additionally, the model enforces a low-rank structure to account for local correlations between labels (e.g., sea and fish).
The optimization is efficiently handled using the Alternating Direction Method of Multipliers (ADMM) framework.
Experimental Evaluation
The authors evaluated TMV-LE against state-of-the-art multi-view learning and label enhancement methods across four diverse datasets: Yeast (biological), Emotions (music), Corel5k, and PASCAL VOC (image recognition). These datasets vary significantly in their number of instances, views, and labels, providing a rigorous testing ground for the algorithm's versatility.
Results and Discussion
As shown in the experimental results, TMV-LE consistently outperformed competing methods:
- Superior Accuracy: TMV-LE achieved the best performance in 81.3% of the experimental configurations across 16 different metrics.
- The Multi-View Advantage: Multi-view strategies consistently outperformed "best single view" approaches, confirming that integrating diverse data sources provides a clearer picture of label ambiguity.
- Ablation Insights: Comparisons with a simplified version (MV-LE) that lacked tensor factorization confirmed that the tensor-based common representation is a critical driver of the model's success.
Conclusion
TMV-LE represents a significant step forward in label enhancement by successfully applying tensor factorization to multi-view data. By capturing high-order relationships and migrating topological structures, it provides a more nuanced and accurate representation of label distributions. Future work may explore end-to-end deep learning architectures to further streamline the representation and enhancement process, as well as strategies for handling missing views in real-world datasets.
Summary updated