BEYOND TWO-TOWER, BACK TO LOCAL: REVISITING CONTEXTGNN FOR TOP-K RECOMMENDATION

ContextGNN is a recent state-of-the-art recommendation framework that extends standard two-tower architectures by explicitly incorporating contextual information. It combines local pair-wise scoring over sampled user contexts with global two-tower scoring across the full item space. Although the original work reports strong results on relational and temporal recommendation tasks, subsequent evidence suggests that ContextGNN is less effective for static link prediction, i.e., classical top-K recommendation. To investigate this limitation, a replication study and methodological re-analysis of ContextGNN are conducted in the top-K recommendation setting, with particular emphasis on the effectiveness of its local-global mechanism. Several architectural variants are explored, including both modifications suggested by the original authors and additional extensions proposed in this work, focusing on four key components: (i) the graph neural network (GNN) backbone, (ii) the local-global fusion strategy, (iii) the contrastive learning objective, and (iv) hard-negative mining. Experiments show that ContextGNN achieves its best ranking performance when recommendations are mainly driven by the sampled local candidate set and locally conditioned scoring. In this setting, the global branch contributes only marginally, and none of the proposed modifications consistently enables a robust interaction between local and global components beyond the standard two-tower paradigm. These results highlight a boundary condition of ContextGNN: without sufficiently informative contextual signals, its advantages over conventional two-tower models largely reduce to local re-ranking over sampled candidates.