Abstract
Partial Label Learning (PLL) addresses scenarios where each training instance is associated with a set of candidate labels, only one of which is correct. This paper introduces CausalPLL+, a framework that integrates Causal Representation Learning (CRL) with PLL to handle instance-dependent label noise. By disentangling latent representations into class-relevant "content" and domain-specific "style," CausalPLL+ achieves superior accuracy and robustness to domain shifts.
Introduction
In real-world applications, obtaining precise ground-truth labels is often difficult or expensive. Instead, we frequently encounter Partial Label Learning (PLL), where annotators provide a set of candidate labels. A significant challenge in this field is Instance-Dependent PLL (IDPLL), where the ambiguity in the label set is correlated with the features of the instance itself. Traditional methods often treat these candidate labels as mere noise. However, this work posits that the structure of these label sets contains critical information about the underlying data-generating process.
Methodology: CausalPLL+
The proposed framework, CausalPLL+, utilizes a Variational Autoencoder (VAE) architecture designed for identifiable causal representation learning.
Latent Disentanglement
The core of the model is the bifurcation of the latent space into two distinct components:
- Content ($z_c$): Captures the invariant features directly tied to the true class label.
- Style ($z_e$): Captures class-irrelevant factors such as lighting, background, or domain-specific textures.
Identifiability and Contrastive Learning
To ensure that the model recovers the true latent factors, the authors employ the theory of identifiable VAEs (iVAE). By conditioning the content prior on the candidate label set, the model can provably distinguish between content and style. Furthermore, a Prior-Based Contrastive Learning module is introduced to sharpen these representations by pulling latent embeddings toward class-specific means.
Experimental Evaluation
The authors evaluated CausalPLL+ on several benchmarks, including FashionMNIST, KMNIST, SVHN, and CIFAR10, under varying levels of label ambiguity ($\tau$).
Quantitative Results
CausalPLL+ consistently outperforms state-of-the-art baselines, particularly as the ambiguity in the partial labels increases. For instance, on CIFAR10 with $\tau=64$, CausalPLL+ achieved an accuracy of 89.66%, significantly higher than PRODEN (86.36%) and RC (74.93%).
Domain Generalization
A key strength of the causal approach is its robustness to domain shifts. In tasks such as USPS $\to$ SVHN, CausalPLL+ maintained a high accuracy (73.15%), whereas traditional PLL methods often failed. This robustness stems from the explicit separation of domain-invariant content from domain-specific style.
Conclusion
CausalPLL+ demonstrates that the ambiguity in partial labels can be leveraged as a structured causal signal rather than just noise. By mathematically disentangling content from style, the framework provides a robust solution for learning from imperfect supervision and generalizing to unseen domains. Future work will focus on scaling these generative causal models to larger datasets like ImageNet.
Summary updated