State-of-the-art sequential recommenders excel at predicting what a user will buy next, yet often fail to predict when. This is due to a flawed assumption: that repurchase intervals follow a single, simple pattern. We empirically demonstrate that the distribution of repeated-purchase intervals (DRPI) is, in fact, a complex mixture: typically a dominant power-law trend overlaid with multiple periodic spikes at weekly, monthly, and other granularities. We formalize this as the Principle of Multi-Granularity Repurchase Cycles. Ignoring this multi-modal reality introduces systematic timing bias, especially for frequently repurchased items. In view of this principle, we propose MgRIA, a novel recommendation paradigm that explicitly models these cycles. MgRIA uses a multi-granularity timestamp embedding to disentangle coexisting periodicities and a distribution-aware scoring mechanism to predict repurchase likelihood over time. Across three real-world datasets (Equity, Tafeng, and Taobao), MgRIA outperforms strong neural baselines such as BERT4Rec, TiSASRec, RepeatNet and DPGN by up to 0.9 absolute points in Recall@K and 0.14 points in Time-MRR@10, with average gains of 0.66 Recall@K and 0.14 Time-MRR@10 on the public Tafeng and Taobao benchmarks. The model also provides interpretability by revealing the specific repurchase cycles driving its predictions. By operationalizing our discovered principle, MgRIA bridges the gap between predicting what and when, while we also discuss practical limitations regarding dataset recency, domain transferability and computational overhead in real-time deployments.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号