Optimal Graph Learning and Nuclear Norm Maximization for Deep Cross-Domain Robust Label Propagation

Wei Wang, Hanyang Li, Ke Shi, Chao Huang, Yang Cao, Cong Wang, Xiaochun Cao

Abstract

Domain adaptation aims to achieve label transfer from a labeled source domain to an unlabeled target domain, where the two domains exhibit different distributions. Existing methods primarily concentrate on designing a feature extractor to learn better domain-invariant features, along with developing an effective classifier for reliable predictions. In this paper, we introduce optimal graph learning to generate a cross-domain graph that effectively connects the two domains, and two domain-specific graphs to capture domain-specific structures. On the one hand, we incorporate the three graphs into the label propagation (LP) classifier to enhance its robustness to distribution difference. On the other hand, we leverage the three graphs to introduce graph embedding losses, promoting the learning of locally discriminative and domain-invariant features. Furthermore, we maximize the nuclear norm of predictions in LP to enhance class diversity, thereby improving its robustness to class imbalance problem. Correspondingly, we develop an efficient algorithm to solve the associated optimization problem. Finally, we integrate the proposed LP and graph embedding losses into a deep neural network, resulting in our proposed deep cross-domain robust LP. Extensive experiments conducted on three cross-domain benchmark datasets demonstrate that our proposed approach could outperform existing state-of-the-art domain adaptation methods.

Introduction

Domain Adaptation (DA) is a fundamental challenge in machine learning, focusing on training models on a labeled "source" domain and deploying them on an unlabeled "target" domain with a different distribution. Two primary obstacles hinder this process: Distribution Shift, where the gap between domains confuses standard classifiers, and Class Imbalance, where majority classes dominate predictions, causing the model to neglect minority classes.

Label Propagation (LP) is a powerful tool for spreading labels based on data similarity, but it is notoriously sensitive to these issues. If the domain gap is too wide, the label "bridge" collapses; if classes are imbalanced, labels from majority classes "flood" the minority ones. This paper introduces Deep Cross-Domain Robust Label Propagation (DCDRLP), a framework designed to make LP resilient to both distribution shifts and class imbalances.

Methodology: Robust Label Propagation

The DCDRLP framework integrates three strategic components into a deep neural network architecture:

  1. Optimal Graph Learning (OGL): Instead of a single similarity graph, the model learns three distinct graphs: a cross-domain graph to connect source and target, and two domain-specific graphs to preserve internal structures. This separation ensures meaningful connections even across wide distribution gaps.
  2. Nuclear Norm Maximization (NNM): To combat class imbalance, the authors maximize the nuclear norm of the prediction matrix. This constraint encourages the model to predict a diverse set of classes, preventing it from collapsing into a "low-rank" state where it only predicts majority classes.
  3. Graph Embedding (GE) Losses: These losses are integrated into a ResNet-based feature extractor to force the learning of domain-invariant features, ensuring that object representations are consistent across both source and target domains.
Overview of the DCDRLP framework, illustrating the integration of OGL for graph embedding and NNM for label propagation.

Experimental Results

The proposed method was evaluated on three benchmark datasets: Office-31, Office-Home, and VisDA-2017. The latter is particularly significant as it features over 200,000 images and severe class imbalance.

DCDRLP consistently outperformed 16 state-of-the-art methods. On VisDA-2017, it achieved an accuracy of 85.6%, a 1.3% improvement over previous benchmarks. Qualitative analysis via t-SNE visualizations confirmed that DCDRLP creates tighter, more distinct clusters, indicating successful domain-invariant feature learning.

Comparison of class-wise accuracy and pseudo-label distribution across different configurations, demonstrating how NNM improves diversity and accuracy for minority classes.

Conclusion and Future Work

DCDRLP demonstrates that treating domains separately during graph construction and enforcing class diversity via rank-based constraints significantly enhances the robustness of Label Propagation. While the approach introduces some computational overhead due to $O(N^2)$ graph operations, the performance gains on imbalanced datasets are substantial. Future research may extend this robust LP approach to more complex tasks such as object detection and semantic segmentation.

Summary updated

Abstract

Domain adaptation aims to achieve label transfer from a labeled source domain to an unlabeled target domain, where the two domains exhibit different distributions. Existing methods primarily concentrate on designing a feature extractor to learn better domain-invariant features, along with developing an effective classifier for reliable predictions. In this paper, we introduce optimal graph learning to generate a cross-domain graph that effectively connects the two domains, and two domain-specific graphs to capture domain-specific structures. On the one hand, we incorporate the three graphs into the label propagation (LP) classifier to enhance its robustness to distribution difference. On the other hand, we leverage the three graphs to introduce graph embedding losses, promoting the learning of locally discriminative and domain-invariant features. Furthermore, we maximize the nuclear norm of predictions in LP to enhance class diversity, thereby improving its robustness to class imbalance problem. Correspondingly, we develop an efficient algorithm to solve the associated optimization problem. Finally, we integrate the proposed LP and graph embedding losses into a deep neural network, resulting in our proposed deep cross-domain robust LP. Extensive experiments conducted on three cross-domain benchmark datasets demonstrate that our proposed approach could outperform existing state-of-the-art domain adaptation methods.
BibTeX
@inproceedings{wang2024optimal,
  author = {Wei Wang and Hanyang Li and Ke Shi and Chao Huang and Yang Cao and Cong Wang and Xiaochun Cao},
  title = {{Optimal Graph Learning and Nuclear Norm Maximization for Deep Cross-Domain Robust Label Propagation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=KMWjMYPKaq},
}
Figure