NI-UDA: Graph Adversarial Domain Adaptation from Non-shared-and-Imbalanced Big Data to Small Imbalanced Applications

Guangyi Xiao, Weiwei Xiang, Huan Liu, Hao Chen, Shun Peng, Jingzhi Guo, Zhiguo Gong

Abstract

We propose a new general Graph Adversarial Domain Adaptation (GADA) based on semantic knowledge reasoning of class structure for solving the problem of unsupervised domain adaptation (UDA) from big data with non-shared and imbalanced classes to specified small and imbalanced applications (NI-UDA). Our goal is to leverage prior hierarchy knowledge to enhance domain adversarial aligned feature representation with graph reasoning. To address the challenges in NI-UDA, we equip adversarial domain adaptation with a Hierarchy Graph Reasoning (HGR) layer and a Source Classifier Filter (SCF). Experiments on two benchmark datasets show our GADA methods consistently improve state-of-the-art adversarial UDA algorithms, with GADA(HGR) improving the F1 score of MDD by 7.19% and GVB-GD by 7.89% on imbalanced source tasks in the Meal300 dataset. The code is available at https://gadatransfer.wixsite.com/gada.

Introduction

Unsupervised Domain Adaptation (UDA) has traditionally focused on scenarios where source and target domains share an identical label space. However, real-world "Big Data" sources are often massive, long-tailed, and contain many categories irrelevant to a specific target application. This creates the Non-shared-and-Imbalanced (NI-UDA) problem, where the source domain contains non-shared classes (classes not present in the target) and both domains suffer from class imbalance.

Illustration of the NI-UDA problem and the GADA alignment process.

Challenges in NI-UDA

The NI-UDA setting presents two primary hurdles:

  1. Negative Transfer: Attempting to align features from source classes that do not exist in the target domain can degrade performance on the target task.
  2. Sparse Class Alignment: In imbalanced datasets, "sparse" or rare classes often lack sufficient samples to form robust feature representations, causing them to be overlooked during standard adversarial alignment.

The GADA Framework

To overcome these challenges, we introduce Graph Adversarial Domain Adaptation (GADA). GADA integrates hierarchical semantic knowledge into the adversarial learning process through two key components:

Hierarchy Graph Reasoning (HGR)

The HGR layer utilizes prior knowledge (such as WordNet hierarchies) to construct a semantic graph of all classes. By treating classes as nodes in a graph rather than independent labels, the model can aggregate local features into hierarchical nodes. This allows rare classes to "borrow" structural information from related common classes (e.g., a "wildcat" sample can leverage features from "house cat" or "feline" nodes). The HGR layer employs Personalized PageRank (PPR) for attention and Graph Convolutional Networks (GCNs) to refine these features.

Source Classifier Filter (SCF)

The SCF acts as a gatekeeper to prevent negative transfer. It filters out low-confidence data from non-shared source classes before they enter the HGR layer. By using a confidence threshold, the model ensures that only relevant source knowledge is used to enhance the feature representations.

The GADA architecture featuring Hierarchy Graph Reasoning (HGR) and Domain Adversarial Alignment.

Experimental Evaluation

We evaluated GADA on two benchmarks: Meal-300 (a food recognition dataset with 300 classes) and Office-Home (modified to simulate NI-UDA conditions).

On the Meal-300 dataset, GADA(HGR) significantly outperformed baseline methods. Specifically, it improved the F1 score of the Maximum Classifier Discrepancy (MDD) algorithm by 7.19% and the Gradually Vanishing Bridge (GVB-GD) algorithm by 7.89% on imbalanced source tasks.

On the Office-Home benchmark, GADA(MDD+HGR) achieved an average accuracy of 76.76% and an F1 score of 75.94%, representing a substantial improvement over the standard MDD baseline (72.96% accuracy, 72.01% F1). These results demonstrate that leveraging hierarchical semantic knowledge is a powerful strategy for handling sparse and non-shared classes in domain adaptation.

Conclusion

GADA provides a robust framework for NI-UDA by combining hierarchical graph reasoning with selective data filtering. By recognizing that classes are semantically related, GADA effectively addresses the data scarcity of rare classes and mitigates the risks of negative transfer from irrelevant source data.

Summary updated

Abstract

We propose a new general Graph Adversarial Domain Adaptation (GADA) based on semantic knowledge reasoning of class structure for solving the problem of unsupervised domain adaptation (UDA) from the big data with non-shared and imbalanced classes to specified small and imbalanced applications (NI-UDA), where non-shared classes mean the label space out of the target domain. Our goal is to leverage priori hierarchy knowledge to enhance domain adversarial aligned feature representation with graph reasoning. In this paper, to address two challenges in NI-UDA, we equip adversarial domain adaptation with Hierarchy Graph Reasoning (HGR) layer and the Source Classifier Filter (SCF). For sparse classes transfer challenge, our HGR layer can aggregate local feature to hierarchy graph nodes by node prediction and enhance domain adversarial aligned feature with hierarchy graph reasoning for sparse classes. Our HGR contributes to learn direct semantic patterns for sparse classes by hierarchy attention in self-attention, non-linear mapping and graph normalization. our SCF is proposed for the challenge of knowledge sharing from non-shared data without negative transfer effect by filtering low-confidence non-shared data in HGR layer. Experiments on two benchmark datasets show our GADA methods consistently improve the state-of-the-art adversarial UDA algorithms, e.g. GADA(HGR) can greatly improve f1 of the MDD by \textbf{7.19\%} and GVB-GD by \textbf{7.89\%} respectively on imbalanced source task in Meal300 dataset. The code is available at https://gadatransfer.wixsite.com/gada.
BibTeX
@misc{xiao2021ni,
  author = {Guangyi Xiao and Weiwei Xiang and Huan Liu and Hao Chen and Shun Peng and Jingzhi Guo and Zhiguo Gong},
  title = {{NI-UDA: Graph Adversarial Domain Adaptation from Non-shared-and-Imbalanced Big Data to Small Imbalanced Applications}},
  year = {2021},
  eprint = {2108.05061},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.05061},
}
Figure