.png){kind=icon}
.png){kind=icon}
Return
PLRA-KG: Pattern-derived latent relation augmentation for knowledge graph-based recommendations
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2026.105109.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge graph (KG)-based recommendation has emerged as an effective solution to data sparsity by incorporating structured semantic information into user and item representations. However, existing KG-based recommendation approaches still face two core challenges: (i) over-reliance on explicit relations while insufficiently capturing latent connections between items, and (ii) feature degradation during high-order propagation, which leads to the dilution of informative signals in learned representations. To this end, we propose PLRA-KG, a pattern-derived latent relation augmentation framework for KG-based recommendation. PLRA-KG uncovers implicit item associations by generating syn-relations and syn-entities from global user preference patterns and relational structures. Specifically, PLRA-KG first performs relation-aware clustering via self-training to group items with strong association patterns. It then introduces a discernment-aware relation-pair selection mechanism to identify relation combinations that significantly influence user decisions. Based on the selected relation-pairs, a pattern-derived syn-relation generation strategy is designed to construct latent relations by jointly modeling global user preferences and relational value interactions. These generated syn-relations and syn-entities are subsequently integrated into the original KG, resulting in an augmented graph that better captures hidden semantic connectivity. Finally, PLRA-KG is optimized in a unified framework that jointly considers recommendation learning, KG embedding, and clustering objectives, enabling seamless integration with various KG-based recommendation models. Extensive experiments on multiple benchmark datasets demonstrate that PLRA-KG consistently improves both recommendation accuracy and diversity, achieving average improvements of approximately 6.64% in Recall, 7.83% in AD, 5.56% in Coverage, and 6.20% reduction in ARP. The source code is accessible at https://github.com/ZZP-RS/PLRA-KG
.
Keywords:
Knowledge graph
Graph augmentation
Latent relation
Recommendation
Journal
I
IF:
6.9
Papers:
337
Citations:
0
Organization
View more organizations