PinEqualizer: Full-funnel cold-start debiasing and exploration¶
Fidelity: core mechanism reproduction. Exercises engagement dropout, content interactions, cohort calibration, an exploration corpus and relevance-scaled UCB on public proxy data.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | |
| First public release | 2026-07-23 (arXiv v1) |
| Author code | No official author code identified (checked 2026-08-09). |
| Local adapter or method key | pinequalizer |
| Local reproduction code | src/auto_research/reproductions/pinequalizer/ |
Paper method and evidence¶
A lack of graph connections and engagement creates a self-reinforcing exposure disadvantage for fresh items. PinEqualizer monitors their share and survival at corpus selection, retrieval, ranking and utility. Models use content/VLM/SID features, individual engagement dropout, content–age–engagement crossings, cohort calibration and fresh/old score regularization. A separate exploration corpus uses Thompson posteriors, real-time impression UCB or Neural Linear UCB. Less observed exposure increases the bonus; search relevance suppresses off-topic exploration.
flowchart LR
A[Content corpus and fresh inventory] --> B[Recall and rank with calibrated scores]
B[Recall and rank with calibrated scores] --> C[Relevance-scaled exploration]
C[Relevance-scaled exploration] --> D[Measure cohort survival]Original paper figure¶
Figure 2 shows the overall method and its main components. From the original paper; copyright remains with the authors.
Core equation¶
The exploration posterior uses model prior \(\alpha\) and strength \(N\), positives \(e\), and exposures \(n\); the neural linear variant uses hidden representation \(\phi(x)\).
Reported results¶
Rollout from 2024 accumulated 350% fresh-content impression growth. Long-running A/B holdouts show underexplored engagement +37%/+13%/+27% on Homefeed/Search/Related Pins. Related Pins component A/B: ranking architecture +8.63% all-fresh and +6.57% underexplored engagement; feature improvements +5.57%/+18.52%. Twenty-eight-day graduate corpus and successful providers grow 41% and 99% year over year.
Local reproduction¶
Matched MovieLens-1M arms use the same user/content features, 100 steps and seeds 42/43/44 with independent test. The treatment adds dropout, crossings, score regularization, cohort Platt calibration and validation-only UCB selection. Overall NDCG@10 drops 16.44%, fresh NDCG@10 increases 448.62%, and underexplored exposure@10 increases 471.02%. All three seeds select UCB alpha=0: extra UCB worsens the validation objective after debiasing pushes exposure. Report the overall precision/exploration tradeoff, not just the fresh metric. Stable metrics.
Fresh is the latest 20% by first observed time, and underexplored falls below the fresh popularity median. These are proxies. No PinCLIP, SID, merchant data, three production surfaces, long-term holdout or real-time Flink features; local NDCG does not establish online engagement.