T-RoPE: Time-Aware Rotary Position Embedding for Sequential Recommendation

What happened

To solve the problem that generative recommendation systems directly adapt large model RoPE without capturing time dimension information, as native RoPE only records interaction order, a research team proposed the time-aware T-RoPE position encoding scheme. Built with modifications including timestamp angle encoding, learnable time coefficients, a multi-scale frequency bank, and non-stationary key rotation, T-RoPE delivered strong performance across multiple public datasets, a 6-billion-scale industrial interaction dataset, and online tests on the Shop app. Its extra computational cost grows linearly with sequence length and attention head dimension, making it suitable for large-scale recommendation model deployment.

Key facts

Research solution

Time-aware rotary position embedding (T-RoPE) for generative sequential recommendation

Public dataset performance

Achieved SOTA across all metrics on 5 public benchmarks: on sparse PixelRec data, HR@10 improved by 78%-130% over the strongest baseline; on Amazon Books, all metrics improved by 8%-12%

Industrial dataset performance

On a 6-billion-interaction e-commerce dataset, improved all metrics by 13%-82% over the HSTU+Time RAB backbone

Online test performance

In A/B tests on the Shop app, conversion rate improved by 0.33% and order volume improved by 0.63%

Deployment cost

Extra computational overhead scales linearly with sequence length and attention head dimension, with full compatibility for native RoPE interfaces

Background

Most large-scale recommendation systems currently use the sequence generation paradigm derived from large models, directly adopting the native rotary position embedding (RoPE) from LLMs. However, native RoPE only records the order of interaction events, and cannot capture key time-related information such as interaction time gaps, cross-scale behavior cycles, and calendar phases. Even when adapted to incorporate timestamps, native RoPE suffers from time translation invariance flaws, making it unable to distinguish seasonal use cases.

Why it matters

For the recommendation industry, this solution bridges the gap between LLM position encoding and the time-specific characteristics of recommendation use cases, proving that time dimension signals can drastically boost recommendation accuracy. For developers, T-RoPE's compatibility with native RoPE interfaces and linearly scaling overhead can significantly reduce adaptation costs. For end users, more accurate recommendations reduce irrelevant content interference and improve browsing and shopping efficiency on e-commerce platforms.