PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation¶
Fidelity: core mechanism. Paper results, local public-data measurements, and omissions are separate.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2605.21752 |
| Organization | TikTok |
| First public release | 2026-05-20 (arXiv v1) |
| Authors' code | The paper did not provide official or author code when checked on 2026-08-24. |
| Adapter / method | pearl-percentile |
| Local reproduction code | src/auto_research/reproductions/pearl_percentile/ |
Original paper¶
Motivation and method¶
Estimate behavior percentiles from multiple comparisons rather than one biased label. Value weighting, bootstrapping, and long-term co-training extend the method to multiple ranking objectives.
flowchart LR
A["Multiple-sample comparisons"] --> B["Unbiased percentile"] --> C["Multi-objective ranking"]Key figure from the paper¶
Figure 1: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.
Core equations¶
Paper results¶
The main online result is Watch Duration +2.10% in a week-long A/B test. Offline tables and online metrics use private or paper-specific protocols. The private paper protocol is not directly comparable with local MovieLens measurements.
Local reproduction¶
Local comparison: The baseline is the shared transition + content scorer. The treatment adds only pearl-percentile on the same users, items, full-catalog candidates, and seed.
MovieLens-100K, 220 users and 360 items, seed 42: NDCG@10 0.0540 → 0.0483 (-10.65%); Hit@10 0.1091 → 0.1000. Validation selects mixture weights only; the test set is not used for tuning.
See the fixed metrics.
Reproduction boundaries¶
The core mechanism runs on MovieLens-100K with the same split, fixed full-catalog candidates, and seed. Private features, the production foundation model, and online traffic are unavailable. The implementation has independent model state and scoring, rather than mapping a paper name to a shared weighted formula. Omitted parts are in the adapter's omitted_core_components. Local changes and paper A/B gains are reported separately.