HICO-GT: Hidden Community Based Tokenized Graph Transformer for Node Classification

ICLR 2026 Conference Submission16978 Authors

Keywords: graph Transformer, node classification, hidden community detection

Abstract: Graph Transformers have been proven to be effective for the node classification task, of which tokenized graph Transformer is one of the most powerful approaches. When constructing tokens, existing methods focus on collecting multi-view node information as the Transformer's input. However, if a type of tokens only includes nodes having relations with a target node from one perspective, it will not provide sufficient evidence for predicting unknown labels. Directly applying self-attention to all tokens may also produce contradictory information as they are selected by distinct rules. Meanwhile, as an emerging concept on graphs, hidden communities refer to those with relatively weaker structures and being obscured by stronger ones. In this paper, inspired by the hidden community clustering method, we design a new multi-view graph Transformer called HICO-GT. We first reconstruct the input graph by merging the original topology and attribute information. Through an iterative process of weakening dominant and hidden communities in turn, we obtain two subgraphs both containing node information of topological relation and attributed similarity, and then generate two token sequences correspondingly. Along with another neighborhood sequence produced on the original graph, they are separately fed into the Transformer and fused afterwards to form the final representations. Experimental results on various datasets verify the performance of the proposed model, surpassing existing graph Transformers.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16978