Retrieval-enhanced, Adaptively Collaborative, and Temporal-aware user behavior comprehension for LLM-based sequential recommendation

The rapid advancement of large language models (LLMs) presents new opportunities for recommender systems. However, LLM-based sequential recommenders often struggle to extract effective user interest signals from long and complex behavior sequences, leading to the sequential behavior incomprehension problem. To address this, we propose Retrieval-enhanced Adaptive Collaborative- and Temporal-aware user behavior comprehension (ReACT), a retrieval-augmented framework that empowers LLMs to better model user interests. ReACT introduces: (i) Temporal Pointwise Mutual Information (TPMI), which integrates temporal and collaborative signals to retrieve relevant historical behaviors; and (ii) Adaptive User Behavior Retrieval (AUBR), which dynamically selects the most informative user behaviors for each recommendation. Extensive experiments on three real-world datasets (MovieLens-1M, Amazon-Book, and MovieLens-100K) demonstrate that ReACT achieves competitive performance while utilizing only approximately 20% of the average user sequence and 5% of the training data. An LLM-as-judger evaluation across three datasets demonstrates that our method achieves the highest selection ratios (78%, 64%, and 64%, respectively), showing that the user behaviors retrieved by ReACT are the most informative and interpretable for LLM-based user behavior comprehension.¹