RoTE: Coarse-to-Fine Multi-Level Rotary Time Embedding for Sequential Recommendation
New 'plug-and-play' module adds time-span awareness to AI recommender models, improving predictions.
A research team from Tsinghua University and the Shenzhen International Graduate School has introduced RoTE (Rotary Time Embedding), a new method designed to significantly improve AI-powered sequential recommendation systems. These systems, used by platforms like Netflix and Amazon to predict a user's next click or purchase, traditionally sequence items by timestamp order but ignore the actual time gaps between interactions. RoTE addresses this by decomposing timestamps into coarse-to-fine granularities (e.g., day, hour, minute) and injecting this multi-level time-span information directly into item embeddings. This allows the AI to better perceive whether a user's actions happened minutes, days, or months apart, capturing both short-term and long-term interest evolution.
RoTE's key advantage is its design as a lightweight, plug-and-play module. It can be seamlessly added to the backbone of existing Transformer-based recommendation models without requiring architectural overhauls. The researchers validated RoTE by applying it to several established models and testing on three public benchmarks. The results were consistently positive, with the enhanced models achieving performance gains of up to 20.11% in NDCG@5, a core metric for ranking accuracy. This demonstrates the method's effectiveness and generality for improving real-world recommendation engines.
The paper, accepted for presentation at the prestigious SIGIR 2026 conference, highlights a critical oversight in current sequential modeling. By making AI models temporally aware of not just order but duration, RoTE provides a straightforward path to more nuanced and accurate user behavior prediction. The code is publicly available, paving the way for immediate experimentation and integration by tech companies aiming to refine their personalization algorithms.
- RoTE models time spans between user interactions at multiple granularities, a feature missing from current sequential recommenders.
- It's a plug-and-play module that boosted benchmark performance by up to 20.11% (NDCG@5) when added to existing models.
- The method is accepted at SIGIR 2026 and code is publicly available for integration into production systems.
Why It Matters
This directly improves the accuracy of recommendation engines used by streaming, e-commerce, and social media platforms, leading to better user engagement.