Multi-source Unsupervised Domain Adaptation on Graphs with Transferability Modeling

TL;DR

This work tackles multi-source unsupervised domain adaptation for graphs, where multiple labeled source graphs are used to classify nodes on an unlabeled target graph. It introduces the Selective Multi-source Adaptation for Graph framework, which learns graph-level and node-level selectors to model transferability, and performs bi-level domain alignment by combining selective optimal transport in the embedding space with weighted knowledge distillation in the label space, trained via a meta-learning loop. Transferability is estimated through self-supervised graph modeling tasks, enabling informed source selection and improved adaptation to the target. Experiments on five graph datasets show state-of-the-art MSUDA performance and highlight the importance of source selection and the dual alignment objectives for robust graph knowledge transfer.

Abstract

In this paper, we tackle a new problem of \textit{multi-source unsupervised domain adaptation (MSUDA) for graphs}, where models trained on annotated source domains need to be transferred to the unsupervised target graph for node classification. Due to the discrepancy in distribution across domains, the key challenge is how to select good source instances and how to adapt the model. Diverse graph structures further complicate this problem, rendering previous MSUDA approaches less effective. In this work, we present the framework Selective Multi-source Adaptation for Graph ({\method}), with a graph-modeling-based domain selector, a sub-graph node selector, and a bi-level alignment objective for the adaptation. Concretely, to facilitate the identification of informative source data, the similarity across graphs is disentangled and measured with the transferability of a graph-modeling task set, and we use it as evidence for source domain selection. A node selector is further incorporated to capture the variation in transferability of nodes within the same source domain. To learn invariant features for adaptation, we align the target domain to selected source data both at the embedding space by minimizing the optimal transport distance and at the classification level by distilling the label function. Modules are explicitly learned to select informative source data and conduct the alignment in virtual training splits with a meta-learning strategy. Experimental results on five graph datasets show the effectiveness of the proposed method.

Figures (5)

• Figure 1: An example of MSUDA for graphs. We want to transfer knowledge from annotated source domains $\{\bm{G}_s^1$, $\bm{G}_s^2$, $\bm{G}_s^3\}$ to $\bm{G}_t$. Regions of the same color denote similar node attributes. It can be observed that source domains and sub-graphs of each domain are of different importance in adapting to $\bm{G}_t$ w.r.t node distributions.
• Figure 2: Adaptation process of . Source importance is estimated with a global graph selector and sub-graph node selector, and is incorporated into this bi-level alignment.
• Figure 3: Influence of the weight of domain alignment loss.
• Figure 4: Influence of the hyper-parameter $\epsilon$, which controls optimal transport in Eq. \ref{['eq:ot_cost']}.
• Figure 5: Analysis on weights assigned to source domains by $g_{\text{sel}}^{\text{global}}$ and $g_{\text{sel}}^{\text{local}}$, by comparing them to the accuracy obtained with single-source domain adaptation.