Deep learning models often struggle with "domain shift"—the performance drop when a model trained on one dataset (source) is applied to a different but related dataset (target). In a paper presented at AAAI-25, researchers introduced HiGDA (Hierarchical Graph of Nodes), a novel framework that leverages graph theory to bridge this gap in Semi-Supervised Domain Adaptation (SSDA).
The Challenge of SSDA
SSDA involves a large labeled source dataset and a target dataset with very few labeled samples. Traditional architectures face distinct hurdles:
- CNNs excel at local textures but often fail to capture the global structure of complex objects.
- Vision Transformers (ViTs) capture global relationships but are prone to incorporating irrelevant background noise.
- Structural Gaps exist because most models process images in isolation, failing to explicitly model category-level relationships across domains during feature extraction.
Methodology: The Hierarchical Approach
HiGDA addresses these issues by treating images and batches as a two-level hierarchical graph structure.
1. Local Graph (LoG)
At the feature level, HiGDA represents an image as a graph of patches. Unlike the all-to-all attention of Transformers, each patch node only connects to its $k$-nearest neighbors in the feature space. This selective connectivity acts as a natural saliency mask, allowing the model to "cluster" the main object and effectively ignore background noise.
2. Global Graph (GoG) and Graph Active Learning
At the category level, HiGDA constructs a graph where each image in a training mini-batch is a node. Edges are established between images of the same category, forcing the model to learn domain-invariant representations. To leverage unlabeled target data, the framework employs Graph Active Learning (GAL), which uses high-confidence pseudo-labels to dynamically integrate unlabeled samples into the global graph.
Key Insights and Qualitative Analysis
A significant advantage of HiGDA is its inherent ability to filter noise without requiring explicit attention maps. By restricting connectivity to feature-similar patches, the local graph naturally links different parts of a semantic object (e.g., the body and tail of a dolphin) while excluding complex backgrounds.
Experimental Results
The researchers evaluated HiGDA on three major benchmarks: Office-Home, DomainNet, and VisDA2017. The results demonstrate a new state-of-the-art for SSDA:
- Office-Home: In 1-shot settings, HiGDA outperformed previous leaders by 5.3%.
- DomainNet: Combined with minimax entropy (MME) techniques, HiGDA achieved 90.5% accuracy in 1-shot tasks, significantly surpassing the ResNet-34 baseline of 56.8%.
- Efficiency: The LoG-Tiny backbone uses only 9.5M parameters—less than half of a standard ResNet-34 (21.6M)—while delivering superior accuracy.
Conclusion
HiGDA represents a shift from "image-wise" to "graph-wise" learning in domain adaptation. By modeling the topological structure of categories across domains and using local graphs to filter noise, it provides a robust and efficient solution for SSDA. Future work will focus on mitigating the impact of noisy pseudo-labels in the global graph to further enhance stability.
Summary updated