Nov 2026· Proceedings of the 35th ACM International Conference on Information and Knowledge Management· pp. 2197-2207· 0 citations· 28 references
TL;DR
DARGE (Dynamic Adaptive Routing of Graph Experts), a novel source-free few-shot cross-graph transfer method for TAGs, dynamically combines experts using a two-level routing strategy and consistently outperforms both unified-encoder methods and multi-expert baselines.
Abstract
Cross-graph transfer on text-attributed graphs (TAGs) is challenging because different graph domains often encode different structure-semantic relationships. Existing unified-encoder methods jointly train a single encoder using multiple source graphs, but such a unified encoder yields only marginal improvements over the best single-source encoder due to knowledge conflict across domains. Recent multi-expert methods alleviate this issue by maintaining multiple experts, yet they also require joint training over all source graphs and do not adapt routing behavior to unseen target graphs. These limit scalability, flexibility, and applicability to unseen target TAGs under two realistic constraints of a source-free environment and few-shot node classification. To overcome these limitations, we propose DARGE (Dynamic Adaptive Routing of Graph Experts), a novel source-free few-shot cross-graph transfer method for TAGs. Instead of jointly training experts and a router, DARGE independently trains source-specific experts aligned via a shared universal LLM-derived codebook, allowing their outputs to remain directly comparable. Given an unseen target graph, DARGE dynamically combines experts using a two-level routing strategy: a label-free router first adapts to the unlabeled target graph structure through self-supervision, and an episode-level calibration further refines soft expert selection using lightweight few-shot class separation signals. Experiments on seven benchmark TAG datasets demonstrate that DARGE consistently outperforms both unified-encoder methods and multi-expert baselines. Further analysis shows that our routing strategy successfully recovers target-relevant expert preferences and scales well to large expert pools through signature-based pre-screening.
Graph learning plays a crucial role in capturing valuable information from both graph structures and node attributes. Despite their success, existing methods struggle to generalize to unseen datasets or classes. To achieve robust generalization, we identify two essential capabilities a graph model should possess: (1) Z...
Yang Liu, Yuan Fang, Deyu Bo et al.· Proceedings of the 35th ACM...· 0 citations
EviGDA is proposed, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert, and shows that EviGDA outperforms state-of-the-art baselines.
Text-attributed graph (TAG) learning aims to leverage both textual attributes and graph structures for node classification. Recent studies have utilized Large Language Models (LLMs) to generate pseudo-labels and rationales for few-label TAG learning. However, existing approaches typically rely solely on the textual att...
Jung-Hyun Kim, Sooho Jung, D. Y. Kang et al.· Proceedings of the 35th ACM...· 0 citations
The success of foundation models in language and vision has motivated Graph Foundation Models (GFMs), which learn transferable knowledge through multi-domain graph pre-training. Existing methods advance GFMs through feature unification and structural learning, and in doing so they rely on the raw graph structure to org...
Yu-Peng Tong, Lianze Shan, Ji-Tao Zhao et al.· 0 citations
Textual-edge graphs represent interactions through both connectivity and relation-specific text. Learning community structure from these graphs requires an encoder that preserves the relevance of individual relations when combining semantic and structural information. We propose Dual-Stage Attention Embedding for Commu...
An emerging approach called cross-domain recommendation transfers knowledge across heterogeneous areas, including books and movies, to improve the user experience. Existing methods often suffer from sparse interactions, semantic gaps across domains, and insufficient explainability. To address these limitations, we prop...
Sharmila Rani Jetti, M. H. M. Krishna Prasad· International Journal of Com...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.