Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking

Hsin-Chiang Chen; Hao Wang; Gedi Zhou; Mingze Gao; Abhishek Kumar; Yiben Yang; Haoyu Wang; Hongjing Xia

cs.IROct 20, 2026

Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking

Authors: Hsin-Chiang Chen, Hao Wang, Gedi Zhou, Mingze Gao, Abhishek Kumar, Yiben Yang, Haoyu Wang, Hongjing Xia,

Organizations: Meta USA

Abstract

A feed ranker scores candidates with a value model, a weighted sum of engagement predictions. Each prediction conflates a baseline, what this viewer, item or request would produce anyway, with the incremental amount by which this candidate exceeds it. The literature mostly removes the baseline. We argue that a multi-signal value model should keep the raw prediction and add the incremental part as a separately weighted term. On Instagram Feed, three debiasing families deployed independently, for viewer habit, item popularity and request context, are one parameterised operator, and a design procedure picks a point in it. Media-type (format) bias goes instead to a debiased label chosen by the same procedure. In a controlled four-arm ablation of deployed Feed terms (neither term, each alone, both), the two-term arm beats both single-term arms at production weights on point estimates in 11 of 13 judged (term, metric) pairs. Of the 26 direct contrasts, 17 are significantly for it and only 1 against it, on impressions. The ablation also suggests that the offline diagnostic should screen signals, not rank them. Such terms have served Instagram Feed production for most of a year. Beyond Feed, Meta, whose apps serve billions of people, has adopted the operator across surfaces, apps and both ranking stages. Its launches there drive sessions and daily active users, the core visitation and retention metrics.

Figures & tables

Panel of Figure 1 — Figure 1. Overview of our methodology. (a) Retrieval and early- and late-stage ranking each score candidates with a value model (VM), a weighted sum of predictions, each prediction mixing a baseline $b$ with the candidate’s incremental excess $\Delta$ . We keep the raw prediction and give $\Delta$ its own weight, with no new model. (b) Viewer and item baselines are precomputed historical rates (the row and column averages of the grid); the request baseline is read from the live candidate pool. (c) One online test ships one point; a separate four-arm ablation tests the two-term form. Three panels: the value model in retrieval and in early- and late-stage ranking with its raw and incremental terms; where the viewer, item and request baselines come from; the design procedure and the four-arm ablation.

Method	baseline $b$	scale $d$	raw at serving
Logit adj. ( Menon et al., 2021 )	class prior	logit $-$ log prior	yes, $-\tau\log\pi$
PDA ( Zhang et al., 2021 )	item popularity	mult. (power)	re-injected
MACR ( Wei et al., 2021 )	viewer, item branch	raw gap	yes, $-c\,b$
PPAC ( Ning et al., 2024 )	pers./global pop.	additive	yes, $+\gamma,\beta$
PAL ( Guo et al., 2019 )	position	mult. factor	dropped
Shallow tower ( Zhao et al., 2019 )	position	additive logit	dropped

Table 1. Published methods in the design space of Eq. 3 . Raw at serving is dropped if the bias factor is removed or its features set to missing, and replaced if a debiased target or score stands in.

signal	quantile $r$	zero %	Var%	$n$
like	$+0.687$	6	22.0	2.90M
comment_tap	$+0.652$	15	17.3	2.89M
reshare_select_recipient	$+0.593$	40	45.4	2.67M
reshare_button_tap	$+0.538$	21	16.6	2.89M
profile_tap	$+0.525$	8	9.3	2.90M
save	$+0.512$	43	17.4	2.67M

Table 2. [Offline] Viewer-axis diagnostics. “Zero” is the share of impressions with a zero feature, on which the clamp binds. Var% is computed against the prediction. Feed time spent has no zero share because its feature is in seconds.

Panel of Figure 2 — Figure 2. [Offline] (a) Within-request distributions, each normalised to its own p10–p90 range. (b) Coefficient of variation of each baseline statistic across simulated requests. The simulation reproduces the ordering the rule needs but understates absolute instability: real requests vary up to 68 $\times$ more. Box plots of within-request prediction shapes, and a bar chart of the coefficient of variation of candidate baseline statistics.

popularity	$r$	duration	$r$
profile_tap	$+0.599$	vc_ratio_80	$-0.418$
reshare_button_tap	$+0.521$	log_time	$+0.216$
like	$+0.456$	comment_ts	$+0.121$
follow	$+0.407$	ts_p75_carousel_v2	$+0.105$
feed_time_spent	$+0.400$	reelsiv_chain_view_time	$+0.084$
save	$+0.352$	ts_p75_video_v2	$-0.017$

Table 3. [Offline] Item-axis diagnostics for popularity (left) and video duration (right). A further 6 duration signals, all with

|r|<0.09

, are omitted.

term	axis	form	approved on
bucketTS-p75 †	item	photo: $\Delta={\color[rgb]{0.1211,0.3047,0.6133}\hat{P}}\big(t{\color[rgb]{0.1328,0.5469,0.5}\,>\,}{\color[rgb]{0.7031,0.3242,0}b_{\text{words}}}\big)$ ; carousel: $\Delta={\color[rgb]{0.1211,0.3047,0.6133}\hat{P}}\big(t{\color[rgb]{0.1328,0.5469,0.5}\,>\,}{\color[rgb]{0.7031,0.3242,0}b_{\text{cards}}}\big)$ ; video: $\Delta={\color[rgb]{0.1211,0.3047,0.6133}\hat{P}}\big(t{\color[rgb]{0.1328,0.5469,0.5}\,>\,}{\color[rgb]{0.7031,0.3242,0}b_{\text{duration}}}\big)$ $\hat{P}$ : predicted probability; $t$ : time spent; $b$ : p75 of $t$ in the post’s bucket of OCR word count, number of cards or duration.	$+0.447\%$ $\ast$ feed time spent
logTS-MR	item	$\Delta={\color[rgb]{0.1211,0.3047,0.6133}p}\;\mathopen{{\color[rgb]{0.1328,0.5469,0.5}\sigma\bigl(}}{\color[rgb]{0.1211,0.3047,0.6133}\hat{t}}-{\color[rgb]{0.7031,0.3242,0}b}\mathclose{{\color[rgb]{0.1328,0.5469,0.5}\bigr)}}$ $p$ : predicted log time spent; $\hat{t}$ : predicted time spent; $b$ : the post’s mean time spent per view.	$+0.665\%$ $\ast$ impressions
seeMore	request	$\Delta=\mathopen{{\color[rgb]{0.1328,0.5469,0.5}\max\bigl(0,}}\;{\color[rgb]{0.1211,0.3047,0.6133}p}-{\color[rgb]{0.7031,0.3242,0}b}\mathclose{{\color[rgb]{0.1328,0.5469,0.5}\bigr)}}$ $p$ : P(viewer taps “See more”); $b$ : mean of $p$ over the request’s candidates; $p$ and $b$ for unconnected candidates only.	$+0.63\%$ $\ast$ unconnected post interested ratio
logTS-TR	item	$\Delta={\color[rgb]{0.1211,0.3047,0.6133}\hat{t}}\;\mathopen{{\color[rgb]{0.1328,0.5469,0.5}\sigma\bigl(}}{\color[rgb]{0.1211,0.3047,0.6133}\hat{t}}-{\color[rgb]{0.7031,0.3242,0}b}\mathclose{{\color[rgb]{0.1328,0.5469,0.5}\bigr)}}$ $\hat{t}$ : predicted time spent; $b$ : the post’s mean time spent per view.	$+0.066\%$ $\ast$ sessions
p10s †	request	$\Delta=\mathopen{{\color[rgb]{0.1328,0.5469,0.5}\max\bigl(0,}}\;{\color[rgb]{0.1211,0.3047,0.6133}p}-{\color[rgb]{0.7031,0.3242,0}b}\mathclose{{\color[rgb]{0.1328,0.5469,0.5}\bigr)}}$ $p$ : P(time spent on the post $>$ 10 s); $b$ : p75 of $p$ over the request’s candidates.	$+0.034\%$ $\ast$ sessions
extShare †	viewer	$\Delta={\color[rgb]{0.1211,0.3047,0.6133}p}\;\mathopen{{\color[rgb]{0.1328,0.5469,0.5}\sigma\bigl(}}\operatorname{logit}{\color[rgb]{0.1211,0.3047,0.6133}p}-\operatorname{logit}{\color[rgb]{0.7031,0.3242,0}b}\mathclose{{\color[rgb]{0.1328,0.5469,0.5}\bigr)}}$ $p$ : P(viewer shares the post outside Instagram); $b$ : the viewer’s own external-share rate.	$+0.030\%$ $\ast$ sessions

Table 4. [Online] The six studied terms, all live. Form : the deployed incremental term

\Delta

, unweighted (Eq. 3 ; for bucketTS-p75, debiased labels), coloured by role: prediction, baseline, the comparison that gates it. Elsewhere a term is named signal

\cdot

axis; TS = time spent, MR/TR = MainRank/TopRank. † Shipped in a bundle; the number is the labels’ own for bucketTS-p75 and the bundle’s for the other two.

\ast

as in Table 5 .

metric	raw	incr	both	raw	incr	JD
	vs. none			both $-$
p10s $\cdot$ req
sessions	$+0.029\%$ $\ast$	$+0.035\%$ $\ast$	$+0.047\%$ $\ast$	$+0.018\%$	$+0.012\%$	✓
feed time spent	$+1.286\%$ $\ast$	$+0.695\%$ $\ast$	$+1.650\%$ $\ast$	$+0.359\%$ $\ast$	$+0.948\%$ $\ast$	✓
logTS-MR $\cdot$ item
sessions	$+0.054\%$ $\ast$	$+0.042\%$ $\ast$	$+0.083\%$ $\ast$	$+0.029\%$ $\ast$	$+0.040\%$ $\ast$	✓

Table 5. [Online] Four-arm ablation. Columns 2–4: each arm relative to none ; columns 5–6: the direct contrasts both

\,-\,

raw and both

\,-\,

incr .

\ast

marks

|\delta|

exceeding the confidence half-width. incr = incremental ; JD = joint dominance. Indented rows are the judged (term, metric) pairs of § 6 .

Panel of Figure 3 — Figure 3. [Offline] Every audited signal on its bias axis. Filled markers carry a term tested online. Their outcomes are in Table 5 , except for item-axis profile tap, whose term is still being iterated. The bucketised labels (ablated together) are an intervention’s output , so the video label’s $r\approx 0$ reflects that intervention working, not a low reading. Offline correlation of every audited signal on its bias axis, with tested signals marked.

stage	term	axis	approved on
LSR	debiased $p_{\text{like}}$	viewer	$+0.083\%$ $\ast$ sessions, $+0.042\%$ $\ast$ DAU
LSR	watch-time gating	item	$+0.049\%$ $\ast$ sessions, $+0.026\%$ $\ast$ DAU
ESR	watch-time gating	item	$+0.029\%$ $\ast$ sessions
ESR	batch $p_{\text{save}}$ champion boost	request	$+0.056\%$ $\ast$ sessions, $+0.032\%$ $\ast$ DAU
ESR	log-time gating	item	$+0.045\%$ $\ast$ sessions
ESR	$p_{\text{like}}$ against viewer like rate	viewer	$+0.075\%$ $\ast$ sessions, $+0.035\%$ $\ast$ DAU

Table 6. [Online] A subset of debiasing launches on the Instagram Reels tab. ESR/LSR = early-/late-stage ranking. Sessions and DAU are each launch’s own online result.

\ast

is defined as in Table 5 .

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 521 papers

MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions
Mar 15, 2026Yuantong Li, Lei Yuan, Zhihao Zheng +27Debiased ML Recommender Systems
Meta AI
DRanker: A Transformer-Based Multi-Task Ranking Model for the Disney+ Homepage
Oct 11, 2026Priya Nirmal Singh Khokher, Daniel Nemirovsky, Benjamin P. Whitesell +4Recommender Systems Learning to Rank
The Walt Disney Company United States
PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation
May 20, 2026Blake Gella, Wei Wu, Yuhao Yin +6Contrastive Learning Learning to Rank
TikTok · San Jose, California, USA · ByteDance

Mar 15, 2026J/K move · Enter open · S save

MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions

Yuantong Li, Lei Yuan, Zhihao Zheng, Weimiao Wu+26

Meta AI

Debiased ML Recommender Systems User Preference Modeling

Open paper AskPDF arXiv

Modern recommendation systems rank candidates by aggregating multiple behavioral signals through a value model. However, many commonly used signals are inherently affected by heterogeneous biases. For example, watch time naturally favors long-form content, loop rate favors short - form content, and comment probability favors videos over images. Such biases introduce two critical issues: (1) value model scores may be systematically misaligned with users' relative preferences - for instance, a seemingly low absolute like probability may represent exceptionally strong interest for a user who rarely engages; and (2) changes in value modeling rules can trigger abrupt and undesirable ecosystem shifts. In this work, we ask a fundamental question: can biased behavioral signals be systematically transformed into unbiased signals, under a user - defined notion of ``unbiasedness'', that are both personalized and adaptive? We propose a general, model-based debiasing (MBD) framework that addresses this challenge by augmenting it with distributional modeling. By conditioning on a flexible subset of features (partial feature set), we explicitly estimate the contextual mean and variance of the engagement distribution for arbitrary cohorts (e.g., specific video lengths or user regions) directly alongside the main prediction. This integration allows the framework to convert biased raw signals into unbiased representations, enabling the construction of higher-level, calibrated signals (such as percentiles or z - scores) suitable for the value model. Importantly, the definition of unbiasedness is flexible and controllable, allowing the system to adapt to different personalization objectives and modeling preferences. Crucially, this is implemented as a lightweight, built-in branch of the existing MTML ranking model, requiring no separate serving infrastructure.

Abstract

Recommender Systems

Explore similar work

MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions

DRanker: A Transformer-Based Multi-Task Ranking Model for the Disney+ Homepage

PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation