KG-ARMM: Residual Structural Fusion for Graph-Augmented Personalized Retrieval
Abstract
Dense multimodal retrieval scales well, but it can miss relational evidence from user histories, item neighborhoods, and domain entities. We present KG-ARMM, a retrieval-compatible Residual Structural Fusion (RSF) framework. A text-and-poster backbone supports approximate nearest-neighbor (ANN) candidate generation; an offline item graph and filtered knowledge graph (KG) supply evidence to a lightweight pairwise re-ranker. The primary configuration gives the learned score limited authority, combining it with normalized structural support, a popularity penalty, and a long-tail bonus. On a temporal MovieLens–The Movie Database (TMDB) benchmark evaluated against the full catalog, the balanced variant improves normalized discounted cumulative gain at 20 (NDCG@20) by 8.3% (0.04273 to 0.04628) and Recall@20 by 7.0% (0.09968 to 0.10666) over the multimodal backbone. A matched full-scale control without graph/KG features or the direct structural term reaches 0.04405 NDCG@20 and 0.10336 Recall@20; complete-model paired gains remain positive with 95% intervals [0.00205, 0.00241] and [0.00285, 0.00375], respectively. KGSupport@20 rises from 0.2704 in that control to 0.3128, with stable catalog behavior. Sparse-history users gain most, and validation-holdout ablations favor the filtered relation bundle. Unrestricted learned-score authority yields higher accuracy but collapses long-tail exposure. These results support structural evidence as a sparse, auditable correction layer over dense multimodal retrieval rather than as an unconstrained replacement for the base ranking signal.