Sparse Causal Discovery with Generative Intervention for Unsupervised Graph Domain Adaptation

Junyu Luo, Yuhao Tang, yiwei Fu, Xiao Luo, Zhizhuo KOU, Zhiping Xiao, Wei Ju, Wentao Zhang, Ming Zhang

TL;DR: SLOGAN boosts Unsupervised Graph Domain Adaptation by using causal discovery, generative intervention, and adaptive calibration to master domain shifts, delivering more robust and accurate graph AI.

Abstract: Unsupervised Graph Domain Adaptation (UGDA) leverages labeled source domain graphs to achieve effective performance in unlabeled target domains despite distribution shifts. However, existing methods often yield suboptimal results due to the entanglement of causal-spurious features and the failure of global alignment strategies. We propose SLOGAN (Sparse Causal Discovery with Generative Intervention), a novel approach that achieves stable graph representation transfer through sparse causal modeling and dynamic intervention mechanisms. Specifically, SLOGAN first constructs a sparse causal graph structure, leveraging mutual information bottleneck constraints to disentangle sparse, stable causal features while compressing domain-dependent spurious correlations through variational inference. To address residual spurious correlations, we innovatively design a generative intervention mechanism that breaks local spurious couplings through cross-domain feature recombination while maintaining causal feature semantic consistency via covariance constraints. Furthermore, to mitigate error accumulation in target domain pseudo-labels, we introduce a category-adaptive dynamic calibration strategy, ensuring stable discriminative learning. Extensive experiments on multiple real-world datasets demonstrate that SLOGAN significantly outperforms existing baselines.

Lay Summary:

When AI learns from one set of graph data (like molecular structures) and tries to apply that knowledge to a new, unlabeled set, it often gets confused. It might focus on **superficial differences** between the datasets (spurious factors) instead of the **true underlying reasons** for a property (causal factors) This is a key challenge in Unsupervised Graph Domain Adaptation (UGDA).

Our method, **SLOGAN**, helps AI make this jump more effectively:

* First, it **disentangles** the essential *causal information* from misleading, domain-specific *spurious details* using sparse causal modeling.
* Then, through a novel **generative intervention**, SLOGAN trains the AI by swapping these *spurious details* between datasets. This forces the AI to rely only on the stable, *causal features*.
* Finally, it carefully uses self-generated labels on the new data with an **adaptive calibration strategy** to ensure reliable learning.

By clearly separating and using *causal patterns* while neutralizing *spurious ones*, SLOGAN significantly boosts the AI's performance and reliability when adapting to new, unlabeled graph data in UGDA. This is a step towards more trustworthy AI for tasks like scientific discovery and network analysis.

Primary Area: General Machine Learning->Transfer, Multitask and Meta-learning

Keywords: Unsupervised Graph Domain Adaptation, Domain Adaptation, Data-efficient Graph Learning

Submission Number: 5785