TisLLM: Temporal Integration-Enhanced Fine-Tuning of Large Language Models for Sequential Recommendation

In recent years, the remarkable versatility of large language models (LLMs) has spurred considerable interest in leveraging their capabilities for recommendation systems. Critically, we argue that the intrinsic aptitude of LLMs for modeling sequential patterns and temporal dynamics renders them uniquely suited for sequential recommendation tasks—a foundational premise explored in depth later in this work. This potential, however, is tempered by significant hurdles: a discernible gap exists between the general competencies of conventional LLMs and the specialized needs of recommendation tasks, and their capacity to uncover complex, latent data interrelationships often proves inadequate, potentially undermining recommendation efficacy. To bridge this gap, our approach centers on adapting LLMs through fine-tuning on dedicated recommendation datasets, enhancing task-specific alignment. Further, we present the temporal Integration Enhanced Fine-Tuning of Large Language Models for Sequential Recommendation (TisLLM) framework. TisLLM specifically targets the deeper excavation of implicit associations within recommendation data streams. Its core mechanism involves partitioning sequential user interaction data using temporally defined sliding windows. These chronologically segmented slices are then aggregated to form enriched contextual representations, which subsequently drive the LLM fine-tuning process. This methodology explicitly strengthens the model’s compatibility with the inherently sequential nature of recommendation scenarios. Rigorous evaluation on benchmark datasets provides robust empirical validation, confirming the effectiveness of the TisLLM framework.