MARS Multi-Resolution Adaptive Routing for Sequential Recommendation
October 7, 2026
MARS mitigates temporal aliasing in long-history recommenders by using recurrent state tracks anchored to different half-lives. This multi-resolution memory architecture allows sparse routing readers to select relevant temporal scales, improving performance on three public datasets by preserving short-term intent and long-term preferences.
HOW THIS AFFECTS YOU
●
builderYou can implement this to improve recommendation engine accuracy for users with long-term interaction histories.
●
researcherThis approach offers a new way to handle multi-scale semantic structures in sequential data.