Taobao's URED recycles high-scored but unexposed candidates: +3.47% discovered categories
_reachsumit · x · 2026-11-06
Alibaba's paper introduces URED (Unexposed Recycling for Efficient Discovery), a training-free recall-stage system that has served Taobao feed recommendation for over six months to tens of millions of DAU.
Key insight: naively boosting under-explored content fails in production because it injects cold, information-poor candidates that rankers systematically over-estimate — no discovery gain, real engagement risk. Items scored highly but never exposed form a pool that is both relevant and reliably predictable, yet industrial pipelines routinely discard this signal.
Approach: URED aggregates these candidates over a 30-day horizon and re-injects them at recall time via a lightweight pipeline (offline storage, distributed aggregation, low-latency serving), adding under 1% latency with a single O(1) lookup per request and no retraining.
Online A/B tests: +3.47% discovered categories, +8.47% content richness, completion video views +2.62%, total views +0.36% (not significant). Ablations show both design pillars are necessary — a one-day window collapses the gains.
More from Research
- ExPert paper inverts perturbation models to trace cell states back to causal causes — anshulkundaje · 2026-11-06
- Follow-up: gradient flow lets residual states encode information for far-future predictions — mlpowered · 2026-11-06
- Gautam Kamath on AI review: fun science experiment, but false positives can mislead people at scale — thegautamkamath · 2026-11-06
- UCLA's FlowPilot uses human preferences to cut sidewalk robot interventions by 40% — zhoubolei · 2026-11-06
- diffTLM: NeurIPS 2026 Spotlight differentiable vocal tract simulator reconstructs shape from voice — sainingxie · 2026-11-06
- AI model learns thousands of words from one child's audio — only with nightly replay, Science reports — jpillowtime · 2026-11-06