Meta-level based recommender system using knowledge graph-based neural collaborative filtering

Abstract

Cold-start remains a major challenge in collaborative filtering, especially when only limited user interactions are available. This study introduces Meta KG-NCF, a hybrid recommender that combines knowledge graph semantics, neural collaborative filtering, and first-order model-agnostic meta-learning (FOMAML). The knowledge graph models books, authors, and publishers using two relation types and TransE embeddings. These embeddings are kept fixed, while FOMAML adapts only the collaborative parameters from five support interactions per new user, reducing the computational cost of second-order optimization. Meta KG-NCF was evaluated on Goodreads, AmazonBook, and Book-Crossing using 80:20 and 90:10 splits under a cold-start protocol with support/query sets and negative sampling. It was compared with five meta-learning baselines: TaNP, PAML, MPML, MetaEDL, and FO-MSAN, under identical CPU-only conditions. Meta KG-NCF reduced MAE and RMSE by approximately 10–11% compared with the strongest baseline, with statistically significant improvements after Holm–Bonferroni correction (p < 10−5), while reducing training time by 11–17%. Its performance on Precision@5, NDCG@5, and MRR@5 was comparable to that of FO-MSAN, except in the cold-start user scenario where Meta KG-NCF achieved MRR@5 = 0.9444, 6.2% points above the closest competitor. Ablation results show that FOMAML is the primary driver of the accuracy gain, while the frozen Knowledge Graph prior contributes an additional consistent tightening on both MAE and RMSE. Comparison against three alternative KG scoring functions (RotatE, ComplEx, TransH) confirms that TransE dominates cold-start user ranking, and a one-factor-at-a-time hyperparameter sensitivity study verifies that the reported results are robust to modest deviations from the paper defaults. Overall, Meta KG-NCF improves cold-start accuracy and efficiency while maintaining competitive ranking quality.