Abstract
Partial label (PL) learning tackles the problem where each training instance is associated with a set of candidate labels that include both the true label and irrelevant noise labels. In this paper, we propose a novel multi-level generative model for partial label learning (MGPLL), which tackles the problem by learning both a label level adversarial generator and a feature level adversarial generator under a bi-directional mapping framework between the label vectors and the data samples. Specifically, MGPLL uses a conditional noise label generation network to model the non-random noise labels and perform label denoising, and uses a multi-class predictor to map the training instances to the denoised label vectors, while a conditional data feature generator is used to form an inverse mapping from the denoised label vectors to data samples. Both the noise label generator and the data feature generator are learned in an adversarial manner to match the observed candidate labels and data features respectively. Extensive experiments are conducted on synthesized and real-world partial label datasets. The proposed approach demonstrates the state-of-the-art performance for partial label learning.
Introduction
In Partial Label (PL) learning, each training example is associated with a set of candidate labels, only one of which is the ground truth. A common real-world example is an image tagged with multiple related objects (e.g., "Computer", "TV", "Monitor"), where the goal is to identify the single true label—a process known as disambiguation.
Most existing PL learning methods rely on the assumption that noise labels are distributed randomly. However, in practice, noise is often label-dependent; for instance, a "Computer" is more likely to be mislabeled as a "Monitor" than a "Tree." Current models often struggle with this non-random noise, either by treating all candidates equally or by using iterative refining processes that can propagate errors over time. To address this, we introduce MGPLL (Multi-level Generative model for Partial Label Learning), a framework that models non-random noise through a multi-level generative adversarial approach.
Methodology
The MGPLL framework consists of five interacting neural networks designed to create a bi-directional mapping between features and labels.
The architecture includes:
- Noise Label Generator ($G_n$): Models the conditional probability of noise labels appearing given a true label.
- Data Feature Generator ($G_x$): Learns to synthesize data features from a given true label.
- Label Discriminator ($D_n$): An adversarial component that distinguishes between generated candidate label sets and observed ones.
- Feature Discriminator ($D_x$): Ensures generated data features match the distribution of real data.
- Prediction Network ($F$): The primary classifier mapping input features to denoised true labels.
By training these components in a minimax adversarial game using the Wasserstein GAN framework, MGPLL establishes a synergy between features and labels. The predictor $F$ maps features to labels, while $G_x$ provides an inverse mapping. This allows the model to discard noise labels that do not align with the underlying data features.
Experiments and Results
MGPLL was evaluated on four synthetic datasets derived from UCI benchmarks (ecoli, vehicle, segment, satimage) and five real-world PL benchmarks: FG-NET, Lost, MSRCv2, BirdSong, and Yahoo! News.
As shown in the results, MGPLL consistently outperforms traditional methods such as PL-KNN, PL-SVM, and recent self-training approaches like SURE. On the MSRCv2 dataset, MGPLL achieved a significant performance gain of 5.2% over the next best method. Similarly, it showed improvements of 3.4% on Yahoo! News and 2.0% on BirdSong.
Ablation studies confirmed that the adversarial label-level loss is the most critical component of the framework. Furthermore, synthetic tests demonstrated that MGPLL remains robust even as the probability of "coupling" (confusing) labels increases, whereas competing models show sharper performance declines.
Conclusion
MGPLL represents a significant step forward in Partial Label Learning by explicitly modeling non-random, label-dependent noise. By leveraging a multi-level generative approach, the model achieves superior disambiguation and state-of-the-art performance across diverse datasets. While the model's complexity and hyperparameter sensitivity present challenges for scaling to extreme classification tasks, its ability to utilize feature-label synergy provides a robust foundation for future research in noisy label environments.
Summary updated