Baseline and Incremental: Debiasing Instagram Feed ranking scores¶
L1 score operator; diagnostic only. No new or trained ranking network or local online lift.
Paper information¶
| Field | Details |
|---|---|
| Paper | Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking |
| Affiliations | Co-first authors Hsin-Chiang Chen and Hao Wang; Meta. |
| First public release | 2026-10-20 00:20:04 UTC |
| Author code | No verified author code repository in paper or arXiv; the production model is not presented as open source. |
| Local key | baseline-incremental; no adapter |
| Local code | src/auto_research/reproductions/baseline_incremental.py |
| Local evidence | Local mechanism metrics |
What the paper changes¶
Engagement reflects user habits, item popularity, request pool and relative candidate quality. Keep the raw prediction while weighting nonlinear baseline-relative increments separately. User/item baselines come from historical features and request baselines from the current pool. This is a scoring change, not causal uplift.
Original paper figure¶
Original figure from the reviewed paper, reused without redrawing.
Core equations¶
Compare probabilities by log-odds and regression quantities in original units. Preserve \(m=p\), \(p-b\) and constant amplitudes, finite sigmoid and exact infinite-slope limits. Explicitly clamp historical features before logit. A negative constant times reverse gate still downweights near the prediction floor; merely negating the prediction-scaled positive gate does not. Infinite-slope gap gates give positive/negative hinges.
Results reported by the authors¶
Section 6 reports a four-arm production test (none/raw/incremental/both), 3% per arm, shared seven days of 12–17 day runs. Table 5 request p10s both vs none increases Feed time spent 1.650%, not additional lift over raw. Single-arm weights were not retuned; no universal optimality. Section 5 reports deployment. Offline correlation is screening, not an online lift ranking; sparse/low-correlation signals remain unresolved.
Local reproduction and limits¶
Synthetic CPU seeds 42/43/44 check rational reference, hinge/sigmoid, floor downweight and ranking with increment removed. request_baseline computes supplied pool statistics, not learned selection. No production data, weight tuning, actual recommendation gains or Evolve.
PYTHONPATH=src python scripts/run_oct21_ranking_components.py \
--component baseline-incremental --device cpu --output runs/baseline-incremental-fresh.json
Full experiment: not reproduced¶
Millions of private log entries, real model scores, ranker and randomized traffic are required. Full study not reproduced; no substitute data, training or retries.