Large‑scale recommender systems are increasingly adopting the sequential generative paradigm of large language models, bringing Transformers and text‑oriented designs such as Rotary Position Embedding (RoPE) into recommendation. Standard RoPE encodes token indices with rotation angles for relative‑position reasoning, but in recommendation the interaction index only records event order, ignoring elapsed time, multi‑scale behavioral cycles, or calendar phases. We revisit this limitation and introduce T‑RoPE, a time‑aware RoPE for sequential generative recommendation. T‑RoPE replaces index‑only rotations with timestamp‑derived angles, adds learnable temporal coefficients, multi‑scale frequency banks, shifted query alignment, and non‑stationary key rotation, thereby breaking the time‑translation invariance of vanilla RoPE while preserving its interface. Across five public benchmarks T‑RoPE achieves the best scores on every metric: HR@10 improves by 78%‑130% on the sparse PixelRec dataset and 8%‑12% on Amazon Books. On an industrial e‑commerce dataset with over 6 B interactions, it outperforms the HSTU + Time RAB backbone by 13%‑82%, with the largest gains from multi‑scale frequencies (+56% NDCG@50) and non‑stationary keys (+4%). An online A/B test in the Shop app yields lifts of +0.33% in conversion rate and +0.63% in order count. We also provide forward and backward algorithms whose overhead is linear in sequence length and head dimension, keeping time‑aware RoPE practical for large generative recommenders.
Review