Aggregate to Adapt: Node-Centric Aggregation for Multi-Source-Free Graph Domain Adaptation

TL;DR

This work tackles unsupervised multi-source-free graph domain adaptation by introducing GraphATA, a node-centric framework that personalizes per-node graph convolutional matrices using multiple source pre-trained models without accessing labeled source data. GraphATA aggregates source-model weights at the node level by computing a local context vector, applying sparsemax-based attention to select relevant sources, and incorporating a global component, yielding node-specific $W_v^l$. The model employs a memory-bank-based pseudo-labeling scheme with neighborhood-aware entropy regularization to drive training, and demonstrates strong, consistent gains over state-of-the-art baselines across node and graph classification benchmarks, including large-scale datasets. The results indicate that fine-grained, context-aware adaptation mitigates negative transfer and highlights the practical viability of MSFDA in privacy-constrained, multi-domain web-scale graphs.

Abstract

Unsupervised graph domain adaptation (UGDA) focuses on transferring knowledge from labeled source graph to unlabeled target graph under domain discrepancies. Most existing UGDA methods are designed to adapt information from a single source domain, which cannot effectively exploit the complementary knowledge from multiple source domains. Furthermore, their assumptions that the labeled source graphs are accessible throughout the training procedure might not be practical due to privacy, regulation, and storage concerns. In this paper, we investigate multi-source-free unsupervised graph domain adaptation, i.e., adapting knowledge from multiple source domains to an unlabeled target domain without utilizing labeled source graphs but relying solely on source pre-trained models. Unlike previous multi-source domain adaptation approaches that aggregate predictions at model level, we introduce a novel model named GraphATA which conducts adaptation at node granularity. Specifically, we parameterize each node with its own graph convolutional matrix by automatically aggregating weight matrices from multiple source models according to its local context, thus realizing dynamic adaptation over graph structured data. We also demonstrate the capability of GraphATA to generalize to both model-centric and layer-centric methods. Comprehensive experiments on various public datasets show that our GraphATA can consistently surpass recent state-of-the-art baselines with different gains.

Figures (6)

• Figure 1: A toy example, where GNN 1 excels in modeling shared interests, whereas GNN 2 is good at capturing geographical proximity. If node B has mixed connection types, simply combining the predictions from GNN 1 and GNN 2 is ineffective, as neither of the source pre-trained GNNs performs well in this scenario.
• Figure 2: An illustrative comparison between existing model-centric methods and our proposed node-centric framework. (a) The target prediction is the weighted combination of source models' predictions. (b) GraphATA performs fine-grained adaptation by considering each node's unique characteristic. The grey box with dash lines shows the personalized convolutional matrix for each node at layer $l$.
• Figure 3: Node homophily ratio distributions.
• Figure 4: Hyper-parameter sensitivity analysis and attention weights visualization.
• Figure 5: Node embedding visualizations in target graph, where colors correspond to different classes in citation networks (C,A$\rightarrow$D).