T-RoPE Adds Time Awareness to Rotary Position Embeddings for Sequential Recommendation

Summary

The paper introduces T-RoPE, a time-aware Rotary Position Embedding for sequential generative recommendation. The authors argue that standard RoPE records interaction order but misses elapsed time, behavioral cycles at different scales, and calendar phase. T-RoPE replaces index-only rotations with timestamp-based angles, learnable temporal coefficients, multiscale frequency banks, shifted query alignment, and non-stationary key rotation while retaining the RoPE interface. The paper proves that standard RoPE, including a timestamp-based variant, remains invariant to time translation and therefore cannot distinguish seasonal contexts; T-RoPE is designed to break that invariance. On five public benchmarks, it achieves the best result on every reported metric and dataset. Relative to the strongest baseline, the method improves HR@10 by 78-130% on the sparse PixelRec dataset and improves metrics by 8-12% on Amazon Books. On an industrial e-commerce dataset containing more than 6 billion interactions, it improves every metric over the HSTU + Time RAB backbone by 13-82%. Ablation results attribute the largest reported gain to multiscale frequencies, which add 56% in NDCG@50, while non-stationary keys add 4%. An online A/B test in the Shop app reports a 0.33% lift in conversion rate and a 0.63% lift in order count. The authors provide forward and backward algorithms with added cost linear in sequence length and head dimension, presenting the method as practical for large generative recommenders.