PaperlayerView exact revision on arXiv

Source-linked AI summary

Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation

Shuo Chang, Yueqi Wang, Zihuan Diao, Ali Montazer, Jiangguo Zhang, Joyneel Misra, Dapeng Hong, Tomer Margolin, Sourabh Bansod, Ningren Han

arXiv:2609.32839v1cs.IRcs.LG

TL;DR

Recommenders infer preference from behavioral signals that can reflect different underlying states and are confounded by factors such as video duration. FLVM models those signals as measurements of latent value using a restricted baseline and routed latent path. On YouTube Shorts, it lifted the primary viewer enjoyment metric by 2.67% in online A/B tests.

  • Problem

    In short-form feeds, observed behavior is an imperfect and confounded measure of user preference, making preference inference difficult.

    Source: §1 · p. 1

  • Method

    FLVM inserts a structured latent measurement layer that combines a restricted baseline over confounding features with routed latent decoders for heterogeneous feedback.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    2.67% lift in the primary viewer enjoyment metric over the production baseline multi-task pre-scorer, with neutral guardrail metrics in the online A/B experiment.

    Source: §1 · p. 2, §4 · p. 5

  • Takeaways & Limitations

    The latent value score can be integrated into an industrial recommender system and yielded offline and online gains.

    Source: §5 · p. 6

    The baseline excludes user-video matching features, so it has no direct path to user-item relevance.

    Source: §3.2 · p. 3

Abstract

from arXiv · show

Recommender systems rely heavily on heterogeneous behavioral feedback to infer user preference. Although abundant, these signals are imperfect measurements: the same observed behavior can arise from different underlying states, such as genuine enjoyment, passive consumption, or inattention. The challenge is especially acute in short-form video, where watch-based signals are strongly affected by measurement confounders such as video duration - the same watch time can imply different levels of preference for videos of different lengths, while ratio-based metrics can systematically favor short videos. As a result, optimizing raw engagement can amplify measurement artifacts rather than improving user value. We propose a Factorized Latent Value Model (FLVM) for measuring user preference from heterogeneous behavioral feedback. The model treats observed behaviors as noisy measurements of a low-dimensional, factorized latent value state and uses structured output heads to model heterogeneous feedback signals. A restricted baseline path captures predictable variation from measurement-confounding features such as video duration, user propensity, and session context, while a routed latent path estimates preference-relevant value advantage. The resulting latent value score can be integrated into an existing recommender system as a ranking feature or ranking score. On YouTube Shorts, a major short-form video platform, this model improves offline metrics and lifts a primary viewer enjoyment metric by 2.67% in online A/B tests.

1 Introduction

Behavioral feedback is an imperfect, confounded measure of preference, especially in sequential short-form feeds. FLVM adds a structured latent measurement layer before ranking and improves offline metrics while lifting a primary viewer enjoyment metric online.

1 Introduction

  • In low-friction sequential feeds, observed behaviors do not reveal preference directly and can reflect enjoyment, passive consumption, or inattention.Watch time is affected by video duration, while completion-ratio objectives can bias toward shorter videos.Source: §1 · pp. 1–2
  • FLVM models heterogeneous behaviors as noisy measurements of a shared latent value state, inserting a measurement layer between raw feedback and ranking optimization.The paper defines the measurement gap as the distance between observed behaviors and the underlying preference being measured.Source: §1 · p. 2
  • FLVM replaces an existing multi-task ranking model in the pre-scoring stage, with its latents serving as primary inputs to the scoring formula.The paper frames preference learning as a measurement problem because behavioral signals are noisy and confounded observations of latent preference.Source: §1 · p. 2
  • The model combines a restricted baseline over confounding features with sparsely routed latent decoders for heterogeneous feedback.The baseline uses duration, session context, and user propensity; feedback heads include watch, engagement, valence, and satisfaction-like labels.Source: §1 · p. 2
  • 2.67% lift in the primary viewer enjoyment metric, with neutral guardrails, and consistent gains on newly uploaded content.Reported for the YouTube Shorts deployment.Source: §1 · p. 2

2 Related Work

Related work distinguishes observed engagement from the preference or satisfaction a recommender seeks to optimize. FLVM connects that measurement motivation to duration debiasing and latent-variable recommendation through an impression-level routed bottleneck paired with a restricted baseline.

Behavior as a noisy measure of preference.

  • FLVM treats impression-level signals jointly as noisy indicators of a structured latent state, rather than optimizing a single survey or retention target.Prior work includes survey, retention, skip, and explicit negative-feedback signals as evidence beyond positive engagement alone.Source: Behavior as a noisy measure of… · p. 2

Watch-time and feedback debiasing.

  • Prior duration-debiasing approaches use counterfactual labels, duration suppression, or reference distributions; FLVM instead learns a residual latent jointly across heterogeneous feedback.Its restricted baseline uses duration, session context, and user propensity.Source: Watch-time and feedback debiasing · p. 2

Variational and multi-task recommendation.

  • Unlike direct multi-task prediction or prior variational recommendation approaches, FLVM learns an impression-level, semantically routed bottleneck paired with a stopped, restricted baseline.The related work discusses variational representations and production systems that inject inferred intent or predict several feedback objectives directly.Source: Variational and multi-task… · pp. 2–3

3 Factorized Latent Value Model

FLVM models each impression’s unobserved, preference-relevant value as a latent state inferred from serving-time features and observed feedback. Its structured heads combine a restricted confounder baseline with a latent path intended to capture residual variation.

3.1 Problem Formulation

  • Latent value is an impression-level construct for a specific user-item pairing in context, inferred rather than directly optimized because it has no label of its own.The model treats observed behaviors as noisy measurements rather than direct preference labels.Source: §3.1 · p. 3
  • The implementation’s baseline features include video duration, user propensity, and session context, while feedback includes watch-time buckets, completion, likes, shares, negative actions, and survey-style feedback.Source: §3.1 · p. 3
  • Each feedback head combines a restricted baseline using confounding features with a latent path using z, then applies a task-specific output transformation.The heads are conditionally independent given z and the baseline features; the split routes predictable measurement-driven variation to the baseline and preference-relevant residual to z.Source: §3.1 · p. 3

3.2 Baseline and Latent Value Advantage

The baseline path captures feedback variation predictable from restricted confounding features, while the latent path estimates a logit-scale value advantage. Stop-gradient training keeps the main feedback objective from updating the baseline.

3.2 Baseline and Latent Value Advantage

  • Because the baseline excludes user-video matching features, it has no direct path to user-item relevance.Source: §3.2 · p. 3
  • The latent path produces a logit-scale value advantage from task-routed components of z, scaled by a learned positive per-signal gain.The positive gain stabilizes the sign conventions used for identifiability.Source: §3.2 · p. 3
  • The baseline and latent contributions form the final decoder logit through the baseline-relative formulation.The paper relates this construction to relative-advantage approaches for watch-time debiasing [8], extending the idea to residuals across heterogeneous feedback signals.Source: §3.2 · p. 3
  • The main feedback objective rewards the latent advantage only for feedback not already explained by the baseline, whose updates come exclusively from the baseline loss.This training scheme encourages confounding effects to be captured by the baseline, yielding a debiased residual.Source: §3.2 · p. 3
  • Without stop-gradient and an auxiliary baseline loss, the decoder could mix baseline features and z, leaving z no reason to encode only residual variation.Source: §3.2 · p. 3

3.3 Factorized Latent Value Variables

FLVM separates latent value into consumption, active response propensity, and engagement valence, then restricts feedback decoders to selected factors. This structure is intended to give latent components more stable semantics.

3.3 Factorized Latent Value Variables

  • FLVM factorizes latent value into consumption-related evidence, active response propensity, and the positive or negative valence of engagement.Valence and active response propensity jointly explain feedback where the same active behavior can have either polarity.Source: §3.3 · p. 3
  • A single unconstrained latent can average across views that differ in consumption, active response, or engagement direction.For example, completion without response, a like, and a dislike can look alike on consumption alone.Source: §3.3 · p. 3
  • Per-signal hard sparse masks restrict each feedback decoder to a small subset of the factorized latent state.Related feedback tasks can share explanatory factors while unrelated tasks are prevented from freely using every dimension.Source: §3.3 · p. 3
  • Consumption-related labels route through z_p, while explicit feedback routes jointly through z_a and z_s.The Figure 1 architecture pairs this routed latent path with a restricted baseline using measurement-confounding features.Source: §3.3 · p. 4, Figure 1 · p. 4
  • Decoder assignment induces more stable latent semantics by limiting which semantic groups each component can influence.The passage contrasts this structure with unconstrained bottlenecks, whose dimensions can rotate or mix without changing prediction quality.Source: §3.3 · p. 4

3.4 Inductive Bias for Semantic Roles

Routing, positive scaling, and feedback-label polarity encourage stable semantic roles for the factorized latents. Training combines feedback prediction, baseline anchoring, and KL regularization.

3.4 Inductive Bias for Semantic Roles

  • Routing masks discourage rotations that mix factors by sending each latent influence to distinct decoder sets.Mixing consumption into decoders for valence-signed feedback would increase loss.Source: §3.4 · p. 4
  • Strictly positive latent-path scaling preserves a factor’s sign across decoders that share it, while valence-signed labels anchor z_s polarity.Watch-time monotonicity and binary engagement intensities play analogous roles for z_p and z_a.Source: §3.4 · p. 4
  • Across three independently trained models, liked and explicitly negative-feedback impressions separate by 1.7–2.0 standard deviations along z_s.Factor signs and ordering were identical across runs, with neutral topline metrics in online A/A testing.Source: §3.4 · p. 4

3.5 Training Objective

  • The training objective has main feedback, auxiliary baseline anchoring, and KL regularization components.The encoder uses a supervised Gaussian variational bottleneck with reparameterized sampling and KL regularization.Source: §3.5 · p. 4
  • The latent distribution is regularized toward a standard normal prior.Source: §3.5 · p. 4

Main feedback loss.

The main feedback loss trains the encoder and routed decoders against post-impression feedback, while the baseline path is trained separately on confounding features. Latent regularization keeps the latent space centered and calibrated.

Main feedback loss.

  • The main feedback loss trains the encoder and routed latent decoders to predict post-impression feedback.Invalid or missing feedback is masked rather than treated as an implicit negative, and this loss does not update the baseline path.Source: Main feedback loss · p. 5

Auxiliary baseline anchoring loss.

3.6 Ranking Integration

At serving time, FLVM scalarizes selected latent factors into a per-item score used as a primary ranking-formula input, replacing the existing multi-task pre-scoring stage.

3.6 Ranking Integration

  • The serving score combines sigmoid(z_s) with softplus(z_p), omitting z_a because it captures active response propensity without preference polarity.The score becomes a primary ranking-formula input and replaces the existing multi-task pre-scoring stage.Source: §3.6 · p. 5

4 Experiments

On YouTube Shorts, FLVM improves the primary viewer enjoyment metric online and shows duration-sliced offline gains on longer videos. Its baseline and latent paths are complementary, and Figure 2 validates the model’s duration-control and factor-separation design choices.

4 Experiments

  • Completion PR-AUC improves by an average of 0.0032 on videos >60s, while FLVM scores lower by an average of 0.0053 on videos <60s across metrics.Overall, FLVM reduces item watch-time RMSE and improves Like and Dislike PR-AUC, while Completion sees a slight drop; Skip is omitted because label definitions differ.Source: §4 · p. 5, Table 1 · p. 5
  • Jointly explaining heterogeneous feedback through a compact latent value state denoises sparse and confounded signals more effectively than per-task heads.This is the authors’ interpretation of the offline results.Source: §4 · p. 5
  • The Full model exceeds either path alone for all signals, with a particularly strong combined effect on Like, which each path predicts poorly by itself.For Completion and Skip, the latent path drives most predictive performance, while the baseline remains necessary for full accuracy.Source: §4 · p. 5, Table 2 · p. 5
  • 2.67% lift in the primary viewer enjoyment metric over the production baseline, with neutral guardrails; cold-start videos also achieve a consistent >2% lift in early engagement and transition to mainstream distribution.The online A/B experiment spanned 14 days at standard experiment traffic allocation.Source: §4 · p. 5
  • Figure 2 shows predicted item watch time increasing with video duration while the served consumption score stays roughly flat; liked and disliked impressions separate most on valence.Their separation is weaker on active engagement and approximately invariant on consumption.Source: Figure 2 · p. 6
  • The learned latent value state robustly generalizes to trailing-edge data distributions by separating genuine preference from structural and statistical noise.Source: §4 · p. 6

5 Conclusion

The conclusion presents FLVM as a factorized latent measurement model with a restricted baseline and sparse, semantically routed decoders. The resulting score integrates into an industrial recommender with offline and online gains and stable semantic roles for its latent factors.

5 Conclusion

  • FLVM treats observed feedback as noisy measurements of a low-dimensional, factorized latent value state.Source: §5 · p. 6
  • Its design combines a baseline restricted to measurement-confounding features, factorized preference variables, and hard sparse routing to semantically relevant feedback signals.Source: §5 · p. 6
  • The latent value score yields meaningful offline and online gains in an industrial recommender, with robust generalization across content lifecycles and data sparsity; learned factors take on stable semantic roles.Source: §5 · p. 6

Related papers

Loading 2609.32839v1…