Enhancing Recommendation with Reliable Multi-profile Alignment and Collaborative-aware Contrastive Learning

Yibin Liu, Jianyu Zhang, Shijian Li

Abstract: Recent studies have explored the integration of Large Language Models (LLMs) into recommender systems to enhance the semantic understanding of users and items. While traditional collaborative filtering approaches primarily rely on interaction histories, LLM-enhanced methods attempt to construct comprehensive profiles by leveraging descriptive metadata and user-generated reviews. The semantic representations of these profiles are then aligned with recommender embeddings to enhance the performance of recommender systems. However, the effectiveness of such approaches heavily depends on the quality of the generated profiles, which face several critical challenges: inaccurate profiles, insufficient information and information gap between semantic representations and recommender embeddings. To tackle these challenges, we propose a novel framework with reliable multi-profile alignment and collaborative-aware contrastive learning. Specifically, we introduce a profile generation method combining Chain-of-Thought(CoT) prompting and self-reflection to address the issue of inaccurate profiles. To alleviate the problem of insufficient information, we introduce an interactive profile construction mechanism that aggregates and summarizes common characteristics from users' and items' neighbors in the user-item graph. To bridge the information gap between semantic representations and recommender embeddings, we propose interactive information fusion(IIF), which aggregates semantic representations from neighbors and employs supervised contrastive learning to guide representation learning. Furthermore, we propose a multi-profile alignment framework that aligns recommender embeddings with both basic profiles and interactive profiles through deduplicated contrastive objectives, facilitating effective semantic-behavioral alignment. Extensive experiments on three public datasets and six base recommenders demonstrate that our method consistently outperforms strong LLM-based baselines, achieving an average improvement of 2.93% in Recall@20 and 2.64% in NDCG@20.

External IDs:doi:10.1145/3746252.3761267