Adversarial Partial Multi-Label Learning with Label Disambiguation

Yan Yan, Yuhong GuoAAAI 2021 · OpenReview

Abstract

Partial multi-label learning (PML), which tackles the problem of learning multi-label prediction models from instances with overcomplete noisy annotations, has recently started gaining attention from the research community. In this paper, we propose a novel adversarial learning model, PML-GAN, under a generalized encoder-decoder framework for partial multi-label learning. The PML-GAN model uses a disambiguation network to identify irrelevant labels and uses a multi-label prediction network to map the training instances to their disambiguated label vectors, while deploying a generative adversarial network as an inverse mapping from label vectors to data samples in the input feature space. The learning of the overall model corresponds to a minimax adversarial game, which enhances the correspondence of input features with the output labels in a bi-directional mapping. Extensive experiments are conducted on both synthetic and real-world partial multi-label datasets, while the proposed model demonstrates the state-of-the-art performance.

Introduction

In many real-world scenarios, such as crowdsourced image tagging, annotators often provide an "overcomplete" set of labels. This means an instance is assigned several candidate labels, some of which are correct (ground-truth) and others that are just noise. This is the Partial Multi-Label Learning (PML) problem. The challenge lies in the fact that the number of true labels is unknown, and they are hidden among irrelevant ones. Traditional multi-label learning fails here because it treats all candidate labels as ground truth, leading the model to learn from noise. Existing PML methods often rely on "label propagation" or confidence scores, but these are prone to cumulative errors—if the initial guess about which labels are noisy is wrong, the whole model suffers.

Methodology

The authors propose PML-GAN, which operates under a generalized encoder-decoder framework. The architecture consists of four interacting sub-networks:

  1. Disambiguation Network ($\tilde{D}$): Predicts the probability that a candidate label is actually noise for a specific instance.
  2. Prediction Network ($F$): The "Encoder." It maps the input features to a disambiguated (cleaned) label vector.
  3. Generation Network ($G$): The "Decoder." It takes a label vector and attempts to reconstruct the original data sample in the feature space.
  4. Discrimination Network ($D$): A standard GAN component that tries to distinguish between real data samples and the "fake" samples reconstructed by the Generator.
The PML-GAN architecture featuring disambiguation, prediction, generation, and discrimination networks.

Bi-directional Consistency

The core idea is bi-directional consistency. If the Prediction network has correctly identified the true labels, the Generation network should be able to use those labels to reconstruct the original input features. By setting up a minimax adversarial game, the model forces the labels and features to be mathematically consistent in both directions, which naturally pushes the noise out of the label set. Unlike two-stage methods that clean labels first and then train a classifier, PML-GAN does both simultaneously, allowing classification performance to provide feedback to the disambiguation process.

Experimental Results

The researchers evaluated PML-GAN on 12 multi-label datasets, including real-world PML datasets like mirflickr and music_style, as well as synthetic datasets transformed from standard benchmarks like yeast and enron by adding irrelevant labels.

Experimental results comparing PML-GAN against state-of-the-art methods across four evaluation metrics.

PML-GAN was compared against five state-of-the-art PML methods (such as PML-NI and PARTICLE) and one baseline (ML-RBF). The model achieved the best results in 40 out of 48 evaluation cases across 12 datasets and 4 metrics. In terms of Average Precision, PML-GAN outperformed the next best method by 4.6% on eurlex_dc, 4.8% on scene, and 3.4% on image. The model consistently outperformed the baseline ML-RBF in 100% of the test cases, proving that its noise-handling mechanism is highly effective.

Conclusion and Future Work

PML-GAN demonstrates that generative adversarial frameworks can act as powerful regularizers for label disambiguation. By ensuring the latent label space corresponds to the underlying data structure, the model achieves state-of-the-art performance in noisy environments. However, like most GAN-based architectures, PML-GAN involves a minimax optimization problem that can be sensitive to hyperparameters and computationally expensive. Future work may address these training complexities and explore adaptations for extremely high-dimensional or sparse feature spaces where reconstruction is more challenging.

Summary updated

Abstract

Partial multi-label learning (PML), which tackles the problem of learning multi-label prediction models from instances with overcomplete noisy annotations, has recently started gaining attention from the research community. In this paper, we propose a novel adversarial learning model, PML-GAN, under a generalized encoder-decoder framework for partial multi-label learning. The PML-GAN model uses a disambiguation network to identify irrelevant labels and uses a multi-label prediction network to map the training instances to their disambiguated label vectors, while deploying a generative adversarial network as an inverse mapping from label vectors to data samples in the input feature space. The learning of the overall model corresponds to a minimax adversarial game, which enhances the correspondence of input features with the output labels in a bi-directional mapping. Extensive experiments are conducted on both synthetic and real-world partial multi-label datasets, while the proposed model demonstrates the state-of-the-art performance.
BibTeX
@inproceedings{yan2021adversarial,
  author = {Yan Yan and Yuhong Guo},
  title = {{Adversarial Partial Multi-Label Learning with Label Disambiguation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=I95DhA1bdHM},
}
Figure