T-RoPE Adds Temporal Awareness to Sequential Recommendation
September 28, 2026
T-RoPE replaces index-only rotation in RoPE with timestamp-based angles and learnable temporal coefficients. This allows Transformer-based recommenders to distinguish seasonal contexts and behavioral cycles that standard RoPE ignores.
HOW THIS AFFECTS YOU
●
researcherThis approach breaks time-translation invariance in sequential models, enabling better modeling of non-stationary temporal data.