Node2Graph: Diagnosing Task Unification in Graph Learning

Graph foundation models aim to unify diverse tasks within a common framework, often by reformulating node and link prediction as subgraph classification. We present a systematic study of this node \(\rightarrow \) subgraph conversion across homophilic and heterophilic benchmarks using both GNN and graph transformer backbones. Our analysis reveals that subgraph classification closely mirrors node classification in homophilic settings but suffers pronounced and growing accuracy losses in heterophilic contexts as the k-hop radius expands. To explain this behavior, we introduce a simple alignment score that quantifies neighborhood label support and reliably predicts the transfer gap. The results expose both the promise and the pitfalls of task unification, offering practical diagnostics and alignment-aware adaptations for more robust and generalizable graph foundation pipelines.