Revisiting Clique and Star Expansions in Hypergraph Representation Learning: Observations, Problems, and Solutions

Hypergraph representation learning has gained increasing attention for modeling higher-order relationships beyond pairwise interactions. Among existing approaches, clique expansion-based (CE-based) and star expansion-based (SE-based) methods are two dominant paradigms, yet their fundamental limitations remain underexplored. In this paper, we analyze CE- and SE-based methods and identify two complementary issues: CE-based methods suffer from over-agglomeration, where node representations in overlapping hyperedges become excessively clustered, while SE-based methods exhibit under-agglomeration, failing to sufficiently aggregate nodes within the same hyperedge. To address these issues, we propose STARGCN, a hypergraph representation learning framework that constructs a bipartite graph via star expansion and employs a graph convolutional network with a tuplewise loss to explicitly enforce appropriate aggregation and separation of node representations. Experiments on seven real-world hypergraph datasets demonstrate that STARGCN consistently and significantly outperforms five state-of-the-art CE- and SE-based methods across all datasets, achieving performance gains of up to 13.2% in accuracy and 10.2% in F1-score over the strongest baseline.