Nexus-rec: a neural explainable unified system for knowledge graph-enhanced recommendations

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommender systems increasingly need to infer nuanced preferences from sparse user–item interactions while also leveraging external semantic structure. This paper introduces NEXUS-Rec (Neural EXplainable Unified System for Recommendations), a unified framework that couples five complementary signals into a single learning pipeline. At the representation level, Contrastive Graph Learning (CGL) provides self-supervised user–item encoders with relation-aware graph augmentation that applies heterogeneous perturbation rates across different knowledge graph relation types, and Knowledge Graph Attention Network (KGAT) enables multi-hop, relation-aware propagation over structured knowledge augmented with bilinear attention for higher-order feature interactions. To temper spurious correlations and compress noisy features, an enhanced Variational Information Bottleneck (VIB) is incorporated. Sequential dynamics are captured through Temporal Attention Flow (TAF), which models time-ordered consumption patterns with an exponential recency decay mechanism that constrains the temporal receptive field, and the resulting component outputs are combined via a rank-aware Gated Fusion mechanism that couples global learnable component gates with candidate-level rank-agreement features, so that the fused score adapts to each user and candidate while the gates expose interpretable component importance. On MovieLens-1M, NEXUS-Rec achieves an NDCG@10 of 0.4006, significantly outperforming every baseline including the strongest, BPR-MF (0.3368, \(+18.96\%\) , \(p < 10^{-24}\) ) and KGAT (0.3341, \(p < 10^{-28}\) ), as well as recent self-supervised graph methods SGL, SimGCL, and LightGCL (all \(p < 10^{-100}\) ). To probe generality, we add a second benchmark (MovieLens-100K), where NEXUS-Rec again attains the best NDCG@10 (0.3712), improving over the strongest external baseline BPR-MF by \(+20.9\%\) . Ablation analyses demonstrate that each component contributes substantively (22–35% drop when removed), with ItemKNN, Content, and KGAT producing the largest individual effects. The learned importance gates provide interpretable component contributions for instance, collaborative filtering receives the highest gate activation (0.317), followed by temporal attention (0.288) and popularity signals (0.159) enabling transparent recommendation explanations. Overall, the framework offers a principled rank-aware gated fusion strategy that achieves superior ranking accuracy across two datasets while exposing interpretable, modular component contributions.