<p>Modern recommender systems are routinely pulled in two directions. On the one hand, they must infer fine-grained preference signals from sparse, noisy interactions; on the other, they are increasingly expected to exploit heterogeneous evidence structured knowledge, semantic content, and behavioral traces without collapsing into brittle, overfit heuristics. With this tension in mind, we propose PRISM-Rec (Progressive Relational Interest and Semantic Modeling Recommender), a unified recommendation framework that couples knowledge graph–enhanced representations with multi-granularity interest modeling through an adaptive fusion strategy. Rather than treating these signals as loosely connected add-ons, PRISM-Rec integrates five tightly coordinated components: (i) a Multi-Relational Graph Attention Network (MR-GAT) that encodes entity interactions under diverse semantic relations, (ii) a Hierarchical Interest Modeling (HIM) module that captures user preferences at multiple abstraction levels, (iii) a Graph-Content Synergy Module (GCSM) that explicitly aligns structural cues with semantic content (though this assumption merits scrutiny), (iv) an Adaptive Boosted Fusion (ABF) network that learns how to weight competing component signals, and (v) a Diversity-Aware Re-ranking (DAR) stage based on Maximal Marginal Relevance to negotiate the relevance–diversity tradeoff in final lists. Empirically, extensive experiments on the full MovieLens-1M dataset (5,950 users, 3,125 items, 567,776 training interactions) show that PRISM-Rec delivers an 11.80% gain in NDCG@10 over the best-performing baseline (GCSM with NDCG@10 = 0.2998), and the improvement is statistically significant (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$p < 0.001$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo><</mo> <mn>0.001</mn> </mrow> </math></EquationSource> </InlineEquation>). Perhaps more revealing is the magnitude of the margin over knowledge graph–centric alternatives: PRISM-Rec significantly exceeds KGAT, yielding +48.11% in NDCG@10. The knowledge graph constructed for this study comprises 589,445 triples spanning five relation types, which supplies a substantially richer substrate for semantic reasoning than interaction-only signals. As noted above, robustness is not implied by aggregate scores alone; accordingly, we further conduct comprehensive stratification analysis across user activity levels (cold/warm/active), item popularity tiers (head/mid/tail), and temporal contexts (early/middle/late), where PRISM-Rec maintains consistent advantages and suggests strong generalizability. Finally, ablation experiments help disentangle where the gains originate: removing GCSM produces the largest performance drop (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$-21.27\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>21.27</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>), underscoring the practical importance of explicitly coupling graph structure with content semantics.</p>

PRISM-Rec: progressive relational interest and semantic modeling for knowledge graph-enhanced recommendation

期刊论文
出版模式：

开放获取
发表日期：

2026年3月10日

查看全文

Muniraja Pasupuleti,
Shashank Mouli Satapathy

摘要

Modern recommender systems are routinely pulled in two directions. On the one hand, they must infer fine-grained preference signals from sparse, noisy interactions; on the other, they are increasingly expected to exploit heterogeneous evidence structured knowledge, semantic content, and behavioral traces without collapsing into brittle, overfit heuristics. With this tension in mind, we propose PRISM-Rec (Progressive Relational Interest and Semantic Modeling Recommender), a unified recommendation framework that couples knowledge graph–enhanced representations with multi-granularity interest modeling through an adaptive fusion strategy. Rather than treating these signals as loosely connected add-ons, PRISM-Rec integrates five tightly coordinated components: (i) a Multi-Relational Graph Attention Network (MR-GAT) that encodes entity interactions under diverse semantic relations, (ii) a Hierarchical Interest Modeling (HIM) module that captures user preferences at multiple abstraction levels, (iii) a Graph-Content Synergy Module (GCSM) that explicitly aligns structural cues with semantic content (though this assumption merits scrutiny), (iv) an Adaptive Boosted Fusion (ABF) network that learns how to weight competing component signals, and (v) a Diversity-Aware Re-ranking (DAR) stage based on Maximal Marginal Relevance to negotiate the relevance–diversity tradeoff in final lists. Empirically, extensive experiments on the full MovieLens-1M dataset (5,950 users, 3,125 items, 567,776 training interactions) show that PRISM-Rec delivers an 11.80% gain in NDCG@10 over the best-performing baseline (GCSM with NDCG@10 = 0.2998), and the improvement is statistically significant ( $p < 0.001$ $p < 0.001$ ). Perhaps more revealing is the magnitude of the margin over knowledge graph–centric alternatives: PRISM-Rec significantly exceeds KGAT, yielding +48.11% in NDCG@10. The knowledge graph constructed for this study comprises 589,445 triples spanning five relation types, which supplies a substantially richer substrate for semantic reasoning than interaction-only signals. As noted above, robustness is not implied by aggregate scores alone; accordingly, we further conduct comprehensive stratification analysis across user activity levels (cold/warm/active), item popularity tiers (head/mid/tail), and temporal contexts (early/middle/late), where PRISM-Rec maintains consistent advantages and suggests strong generalizability. Finally, ablation experiments help disentangle where the gains originate: removing GCSM produces the largest performance drop ( $-21.27\%$ $- 21.27 %$ ), underscoring the practical importance of explicitly coupling graph structure with content semantics.