简体中文 · English

ClockRoPE: Modeling periodic user habits with random Fourier rotations¶

Fidelity: core-mechanism reproduction. Runs periodic Gaussian Fourier-kernel modulation of sequence attention. Public data lacks YouTube's hourly behavioral context, so results verify the mechanism and its limits only.

Paper information¶

Field	Details
Paper	arXiv 2607.26369
Institution	YouTube / Google DeepMind
First public release	2026-07-29 (arXiv v1)
Author code	No author-released code found as of 2026-08-24.
Local adapter or method key	`clockrope`
Local reproduction code	`src/auto_research/reproductions/clockrope/`

Original paper summary¶

Background and main changes¶

Standard RoPE's log-linear frequencies favor weaker correlation with greater distance, but user behavior often follows daily and weekly cycles. ClockRoPE samples rotation frequencies from the Fourier transform of a target periodic kernel, strengthening the query-key dot product around the same hour or weekday. It can use half the channels while standard RoPE uses the other half.

flowchart LR
  H[Timestamped behavior sequence] --> QK[Query / key]
  T[Daily and weekly priors] --> F[Fourier frequency sampling]
  F --> R[ClockRoPE random rotation]
  QK --> R
  R --> A[Periodically modulated attention]
  A --> G[Generative retrieval]

Key figure from the paper¶

Figure 3: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

For a positive-definite kernel \(f\), random Fourier rotations satisfy

\[ \mathbb E_{\xi\sim\tau}\left[(R_\xi(p_m)q_m)^\top(R_\xi(p_n)k_n)\right] =q_m^\top k_n f(p_m-p_n). \]

The local implementation uses daily and weekly periodic Gaussian kernels:

\[ f_T(\Delta t)=\exp\{\kappa[\cos(2\pi\Delta t/T)-1]\},\quad T\in\{24,168\}. \]

Offline and online results¶

In a 14-day online A/B test, each variant received 1% of traffic. RoPE+ClockRoPE improved Engagement and Valued Engagement by +0.08% each. After production deployment, Valued Engagement Time rose +0.08% and TPU serving cost fell -0.63%, with neutral latency.

Local reproduction¶

Local comparison: The baseline is a RoPE-style recency ranker without periodic modulation. The experimental arm adds daily and weekly ClockRoPE kernels. Relative NDCG@10 changes by -2.30%.

In the one-seed MovieLens-100K mechanism experiment, Hit@10/NDCG@10 are 0.1091/0.0540 for the RoPE-style baseline and 0.1091/0.0528 for ClockRoPE. Relative NDCG changes by -2.30%. This negative result shows that periodic priors do not automatically help without real hourly or weekly routines.

auto-research reproduce --paper clockrope --dataset-dir data --seed 42

Stable metrics: metrics/movielens-100k-seed42.json.

Scope¶

Public MovieLens provides sparse rating timestamps. The local run constructs periodic positions from deterministic event intervals, rather than reproducing YouTube's private logs, generative retrieval model, TPU kernel, or online serving.