Residual Neighborhood Information in Deep Collaborative Filtering: Evidence from Post-Training Score Augmentation

Collaborative filtering (CF) recommenders are commonly evaluated in isolation, which does not reveal whether a trained model has fully captured the predictive information available from local user neighborhoods. We investigate this question through a paired post-training intervention that augments each frozen recommender with the same external neighbor-consumption signal derived from the uniform average of raw-cosine neighbors’ interactions. Across eight VAE-based, graph-based, and self-supervised backbones on three datasets of contrasting density, the augmentation improves both Recall@30 and NDCG@30 in 20 of the 24 backbone–dataset settings. Averaging the five-run means equally across these settings, Recall@30 increases from 0.223 to 0.254 and NDCG@30 from 0.190 to 0.222, although the gains are not universal. Neighbor-only and similarity-weighted neighborhood baselines generally underperform the fusion, indicating complementarity in the aggregate. Finally, an inference-time interpolation from uniform aggregation to learned attention yields only modest, dataset-dependent accuracy gains near the uniform setting, while larger attention contributions degrade both ranking accuracy and consistency with the fixed reference neighborhood. Overall, a simple model-agnostic neighbor-evidence term provides an effective post-training augmentation, and learned weighting is not required to obtain its primary benefit.