Rethinking Graph Domain Adaptation: A Spectral Contrastive Perspective

TL;DR

Rethinking Graph Domain Adaptation addresses the challenge of domain shifts in graph-structured data for transfer learning. The authors propose FracNet, which decomposes graphs into low- and high-frequency components via graph spectral analysis and couples it with a spectral contrastive framework consisting of SMMI and FMMD. They provide theoretical arguments showing tighter alignment bounds and demonstrate empirical gains across three benchmark datasets, including Mutagenicity, NCI1, and PROTEINS, with challenging transfer tasks seeing substantial improvement. The approach mitigates the blurring of decision boundaries common in domain adaptation by using a contrastive objective and a frequency-aware kernel to align distributions across frequencies. The work advances graph domain adaptation by offering a principled, spectral-aware method with solid theory and strong empirical performance.

Abstract

Graph neural networks (GNNs) have achieved remarkable success in various domains, yet they often struggle with domain adaptation due to significant structural distribution shifts and insufficient exploration of transferable patterns. One of the main reasons behind this is that traditional approaches do not treat global and local patterns discriminatingly so that some local details in the graph may be violated after multi-layer GNN. Our key insight is that domain shifts can be better understood through spectral analysis, where low-frequency components often encode domain-invariant global patterns, and high-frequency components capture domain-specific local details. As such, we propose FracNet (\underline{\textbf{Fr}}equency \underline{\textbf{A}}ware \underline{\textbf{C}}ontrastive Graph \underline{\textbf{Net}}work) with two synergic modules to decompose the original graph into high-frequency and low-frequency components and perform frequency-aware domain adaption. Moreover, the blurring boundary problem of domain adaptation is improved by integrating with a contrastive learning framework. Besides the practical implication, we also provide rigorous theoretical proof to demonstrate the superiority of FracNet. Extensive experiments further demonstrate significant improvements over state-of-the-art approaches.

Figures (7)

• Figure 1: Molecular topological structures of four compounds.
• Figure 2: Molecular properties of four compounds.
• Figure 3: Framework overview of FracNet. The model decomposes source and target graphs into high/low frequency components via Fourier Transform, followed by dual-stream processing with SMMI for contrastive learning and FMMD for domain alignment.
• Figure 4: Energy trajectories in frequency domain on three datasets.
• Figure 5: Pairwise spectral differences, decomposed into low-frequency (blue) and high-frequency (pink) components.