TRUST framework redefines temporal session-based recommendation with item-calibrated intervals

Temporal signals have long been used in session-based recommendation to infer user interest, but a new paper from Linjiang Guo, Nitin Bisht, Shiqing Wu, Yifan Yin, and Guandong Xu challenges a core assumption. Most existing temporal session-based recommenders treat time intervals as absolute—implicitly assuming that a 30-minute gap between clicks means the same thing for all items. The authors empirically show this is false: each item has its own interval distribution, so an interval should be interpreted relative to the item it belongs to. Based on this observation, they propose TRUST (Item-Calibrated Interval Evidence for Temporal Session-Based Recommendation). TRUST introduces a score function that evaluates each observed interval against the empirical distribution of the corresponding item. This score guides three key components: global neighbor sampling, session graph encoding, and final interest aggregation. The result is a recommendation system that adapts its temporal understanding per item, rather than applying a one-size-fits-all decay.

Experiments on public datasets demonstrate that TRUST consistently outperforms both temporal and non-temporal baselines. Moreover, the proposed scoring function works as a model-agnostic plug-in—integrating it into existing temporal recommenders yields immediate improvements. Component-wise ablations reveal that calibrating temporal signals within each module (neighbor sampling, graph encoding, interest aggregation) consistently beats removing the module entirely. This suggests the real value lies in the calibration, not just the architectural choice. For practitioners building session-based recommenders, TRUST offers a principled way to handle the nuanced role of time, potentially improving personalization in e-commerce, streaming, and content platforms where session patterns vary widely across items. The paper is available on arXiv under identifier 2606.27214.