Metadata-Gated Dual-Channel Graph Neural Networks for Fine-Grained Recommendation Personalization

Abstract

Graph neural networks (GNNs) have become a dominant paradigm for recommendation, yet most existing metadata-aware variants treat side information as an auxiliary feature that is simply concatenated onto node embeddings before standard message passing. This design implicitly assumes that metadata is always available, equally reliable across users and items, and equally relevant at every propagation step assumption that break down under sparse profiles, incomplete attributes, and cold-start conditions. This paper addresses that gap by proposing a Metadata-Gated Dual-Channel Graph Neural Network (MG-DGNN) that separates structural (interaction-based) and semantic (metadata-based) propagation into two explicit channels and fuses them through a learnable, per-node gating function conditioned on metadata confidence, trained jointly with a hybrid objective combining Bayesian Personalized Ranking (BPR) loss with a metadata-alignment contrastive term. We position MG-DGNN relative to GraphSAGE, Graph Attention Networks (GAT), LightGCN, and knowledge-graph attention models (KGAT), and evaluate it on MovieLens-1M (6,040 users, 3,416 items after 5-core filtering) under a leave-one-out protocol with HR@K, NDCG@K, Precision@K, MAP, MRR, coverage, a four-way ablation study, sensitivity analysis over graph depth, embedding dimension, and metadata neighbourhood size, a stratified cold-start evaluation, and a metadata-masking robustness test. The gated dual-channel design significantly outperforms single-channel structural propagation (paired t-test, p < 0.01 across all reported ranking metrics) and shows a clear benefit specifically for cold-start users (HR@10 0.467 vs. 0.367 for matrix factorization on a 30-user cold stratum), consistent with its design intent. It does not, however, show a statistically significant aggregate improvement over plain matrix factorization at the training budget used here, and the ablation study does not confirm a measurable benefit from the metadata-alignment contrastive term in a single run. We report these results without adjustment, including the ones that do not favor the proposed method, and identify the specific further evidence multi-seed ablation runs, a second dataset, and evaluation beyond the roughly 9,500-node scale tested here needed to establish the framework's practical benefit more conclusively.