Introduction
In the field of machine learning, Partial-Label Learning (PLL) addresses scenarios where training data is "inexact." Instead of a single ground-truth label, each instance is associated with a set of candidate labels, only one of which is correct. For instance, an image of a "dog" might be labeled with the set {dog, cat, fox}. Traditionally, PLL research has focused on "label disambiguation"—the process of identifying the true label within the candidate set to then train a standard classifier.
However, most existing PLL methods utilize a uniform feature space for all classes. This paper argues that such an approach is suboptimal. In many domains, specific features are more discriminative for certain classes than others (e.g., texture for "silk" vs. color for "ocean"). While Label-Specific Features (LSF) have proven effective in multi-label learning, they were previously difficult to apply to PLL because LSF typically requires known ground-truth labels to group data for feature extraction. This paper introduces UCL (UnCertainty-aware Label-specific features), a framework designed to overcome this hurdle.
The UCL Framework
The UCL framework shifts the focus from label space cleaning to feature space enhancement through a five-step process:
- Graph-Based Label Enhancement: To compensate for missing ground-truth labels, UCL uses a graph-based approach to generate "soft" pseudo-labels (probabilities) based on instance similarity in the original feature space.
- Uncertainty-Aware Partitioning: This is the core innovation. For each class, the data is partitioned into three regions based on confidence thresholds:
- Positive ($P$): High confidence the label is correct.
- Negative ($N$): High confidence the label is incorrect.
- Uncertain ($U$): Confusing samples that might otherwise introduce noise if forced into $P$ or $N$.
- Spectral Clustering: The framework performs spectral clustering on these three sets to identify "prototypes" (cluster centers) representing the essence of each region.
- Feature Transformation: Original features are transformed into a new space based on the distance of an instance to these class-specific prototypes.
- Classification: A family of binary classifiers is trained on these specialized, enhanced features.
Experimental Evaluation
The researchers evaluated UCL against eight state-of-the-art baselines using both synthetic and real-world datasets:
- Synthetic Datasets: Five UCI datasets (zoo, glass, yeast, tmc2007, sport) were adapted for PLL by adding false positive labels with varying degrees of difficulty.
- Real-World Datasets: Five diverse datasets were used, including Lost and Soccer Player (face naming), MSRCv2 and Mirflickr (image classification), and BirdSong (audio classification).
Key Results and Analysis
UCL demonstrated decisive superiority over existing methods. On real-world datasets, UCL outperformed baselines in 85% to 95% of tested cases. A significant finding was the effectiveness of the "Uncertain" buffer; ablation studies showed that the three-way split (Positive, Negative, Uncertain) was significantly more effective than a standard binary split, as it prevents ambiguous data from degrading the feature generation process.
Furthermore, the framework proved robust to varying hyperparameter settings. Analysis of the positive and negative thresholds ($\tau_h, \tau_l$) and other parameters showed that UCL maintains stable performance across a wide range of values, confirming its practical applicability.
Conclusion
UCL demonstrates that enhancing the input representation through label-specific features can be more effective than simply attempting to clean output labels in PLL. By explicitly handling uncertainty during feature extraction, UCL provides a robust and versatile strategy that can be integrated with various existing PLL algorithms. Future work may involve leveraging deep neural networks to automate the generation of these label-specific features, potentially moving toward end-to-end deep learning architectures.
Summary updated