Answers are grounded in this paper and cite their sources.
Sequential recommender systems often suffer from temporal sparsity caused by irregularly spaced user interactions, which makes modeling evolving preferences difficult. Most existing approaches treat user histories as uniformly ordered sequences and overlook the importance of time intervals between actions. To overcome this limitation, we propose TiLLM-Rec, a Temporal-Interval-Aware Large Language Model for sequential recommendation. TiLLM-Rec introduces a temporal interval encoder that directly integrates irregular interaction gaps into the LLM attention mechanism. By jointly modeling item semantics and temporal dynamics, the model adaptively adjusts attention weights to better capture both what users prefer and when interactions occur. Experiments on multiple benchmark datasets show that TiLLM-Rec consistently outperforms state-of-the-art methods, particularly under sparse and irregular interaction settings.
No takes yet. Share an insight, caveat, or question.
Hosseinzadeh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: