ChronoID adds explicit temporal signals to generative recommendation

Generative recommendation models rely on semantic IDs to represent items and user preferences, but these IDs have a fundamental flaw: they are entirely time-agnostic. Interactions occurring under vastly different temporal contexts—such as browsing habits before and after a major event—get mashed into identical representations. This assumes user intent and item semantics are stationary, which clashes with real-world dynamics like seasonal trends, evolving tastes, or session rhythms. A team of researchers from University of Rochester, IBM Research, and the University of Hong Kong tackled this gap with ChronoID, a unified framework that explicitly incorporates time into semantic ID learning.

ChronoID systematically explores the design space along three orthogonal dimensions of temporal signals: granularity (e.g., hour vs. week), encoding scheme (e.g., absolute vs. relative), and fusion mechanism (how time integrates with item semantics). Beyond proposing a method, the team also contribute a new benchmark for time-explicit generative recommendation, enabling standardized evaluation. Their experiments reveal which temporal infusion strategies work best, how to design the model architecture, and where the gains come from—often from disentangling temporal patterns from semantic similarity. This work directly addresses a core limitation in current recommender systems, promising more context-aware and adaptive recommendations.