Semantic-collaborative restricted Boltzmann machine (SCRBM) for recommender systems: architecture, graph-enhanced extension, and cross-domain evaluation

Micro and Nano Systems Letters · Springer Nature · 2026-11-16 · DOI: 10.1007/s12652-026-05138-8

Аннотация

Two problems keep recurring in deployed recommendation services: sparse interaction data makes purely collaborative approaches unreliable, while graph-based propagation methods, despite their structural advantages, remain indifferent to what items actually mean. We propose the Semantic-Collaborative Restricted Boltzmann Machine (SCRBM), which fuses an energy-based collaborative pathway with a Sentence-BERT (SBERT) semantic branch via a learnable convex blend, and its graph-enhanced extension, SCRBM+, that adds density-adaptive four-way gating over collaborative, semantic, graph-smoothed, and hybrid signals. Both models are trained with a rank-discounted Bayesian Personalised Ranking (BPR) objective that concentrates the loss signal on the hardest unobserved items, weighted by Discounted Cumulative Gain (DCG)-style discounts so ranking errors near the top of the list are penalised most heavily. On MovieLens-1 M (leave-one-out (LOO), 100-negative protocol), SCRBM achieves Hit Rate(HR)@10 = 0.7367 ± 0.0051 and Normalized Discounted Cumulative Gain(NDCG)@10 = 0.5513 ± 0.0021 across five seeds, surpassing LightGCN by 5.7% on HR@10 (paired t-test p < 0.0001, Cohen’s d = 9.06). SCRBM+ gains + 34.9% HR@10 over LightGCN on MovieLens-10 M and + 15.9% on Yelp2020. On Yelp2020, SCRBM+ achieves HR@10 = 0.5768, outperforming LightGCL (ICLR 2023) by + 5.4% under a controlled comparison. Cold-start users with fewer than 20 interactions see up to 87.1% HR@10 improvement over Neural Collaborative Filtering (NCF). Ablation on MovieLens-1M confirms the semantic branch as the dominant contributor (+ 5.22% HR@10), with LayerNorm and item bias providing complementary precision gains; progressive ablation on Yelp2020 confirms the density-adaptive gate as the critical SCRBM+ component (+ 197% HR@10 over the no-gate variant).

Читать у издателя →

Похожие статьи