Multi-LLM semantic fusion with uncertainty-aware GCNs for personalized recommendation

Abstract

Recommender systems (RSs) must effectively integrate structural interaction patterns with rich semantic information while remaining efficient and reliable. We present TriFuse-LightGCN, an uncertainty-aware fusion framework that combines graph-based collaborative filtering with semantic representations produced by multiple locally hosted large language models (LLMs). The approach incorporates cross-view structural–semantic alignment, tri-consistency regularization across LLM priors, and an adaptive gating mechanism that adjusts semantic and structural contributions according to estimated uncertainty during message passing. We evaluate the framework on four benchmark datasets from e-commerce, review-based, and entertainment domains using standard accuracy metrics (Recall@K, NDCG@K, Hit Rate, and MRR) as well as complementary measures reflecting diversity, novelty, catalog coverage, calibration, and robustness under perturbation. We additionally assess efficiency in terms of convergence behavior, inference latency, and memory usage. Across all datasets, TriFuse-LightGCN demonstrates consistent improvements over strong graph-based, contrastive, sequential, and generative baselines. The model provides notable gains in cold-start and long-tail settings, enhances diversity- and coverage-oriented recommendations, and achieves better calibration and robustness to textual and graph noise, all while maintaining inference efficiency comparable to standard GCN architectures. These results indicate that uncertainty-aware fusion of multi-LLM semantic cues with graph-based representations offers a scalable and effective solution for modern personalized recommendation.