Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Guiding the Recommender: Information-Aware Auto-Bidding for Content Promotion

Yumou Liu, Zhenzhe Zheng, Jiang Rong, Yao Hu, Fan Wu, Guihai Chen

TL;DR

This work tackles the cold-start challenge in content promotion by showing that naive impression-maximization can degrade high-quality content. It proposes an information-aware framework that jointly optimizes short-term engagement and long-term model improvement using a tractable gradient-coverage surrogate tied to Fisher-information concepts. A two-stage bidding algorithm with dynamic budget pacing and a confidence-gated gradient estimator (including a zeroth-order variant for black-box models) delivers theoretical guarantees on submodularity, sublinear regret, and budget feasibility, with extensive offline evaluations demonstrating improved AUC and LogLoss and robust budget adherence. The approach reframes paid promotion as principled data acquisition that strengthens recommendation systems and enhances long-term organic outcomes beyond traditional promotion strategies.

Abstract

Modern content platforms offer paid promotion to mitigate cold start by allocating exposure via auctions. Our empirical analysis reveals a counterintuitive flaw in this paradigm: while promotion rescues low-to-medium quality content, it can harm high-quality content by forcing exposure to suboptimal audiences, polluting engagement signals and downgrading future recommendation. We recast content promotion as a dual-objective optimization that balances short-term value acquisition with long-term model improvement. To make this tractable at bid time in content promotion, we introduce a decomposable surrogate objective, gradient coverage, and establish its formal connection to Fisher Information and optimal experimental design. We design a two-stage auto-bidding algorithm based on Lagrange duality that dynamically paces budget through a shadow price and optimizes impression-level bids using per-impression marginal utilities. To address missing labels at bid time, we propose a confidence-gated gradient heuristic, paired with a zeroth-order variant for black-box models that reliably estimates learning signals in real time. We provide theoretical guarantees, proving monotone submodularity of the composite objective, sublinear regret in online auction, and budget feasibility. Extensive offline experiments on synthetic and real-world datasets validate the framework: it outperforms baselines, achieves superior final AUC/LogLoss, adheres closely to budget targets, and remains effective when gradients are approximated zeroth-order. These results show that strategic, information-aware promotion can improve long-term model performance and organic outcomes beyond naive impression-maximization strategies.

Guiding the Recommender: Information-Aware Auto-Bidding for Content Promotion

TL;DR

This work tackles the cold-start challenge in content promotion by showing that naive impression-maximization can degrade high-quality content. It proposes an information-aware framework that jointly optimizes short-term engagement and long-term model improvement using a tractable gradient-coverage surrogate tied to Fisher-information concepts. A two-stage bidding algorithm with dynamic budget pacing and a confidence-gated gradient estimator (including a zeroth-order variant for black-box models) delivers theoretical guarantees on submodularity, sublinear regret, and budget feasibility, with extensive offline evaluations demonstrating improved AUC and LogLoss and robust budget adherence. The approach reframes paid promotion as principled data acquisition that strengthens recommendation systems and enhances long-term organic outcomes beyond traditional promotion strategies.

Abstract

Modern content platforms offer paid promotion to mitigate cold start by allocating exposure via auctions. Our empirical analysis reveals a counterintuitive flaw in this paradigm: while promotion rescues low-to-medium quality content, it can harm high-quality content by forcing exposure to suboptimal audiences, polluting engagement signals and downgrading future recommendation. We recast content promotion as a dual-objective optimization that balances short-term value acquisition with long-term model improvement. To make this tractable at bid time in content promotion, we introduce a decomposable surrogate objective, gradient coverage, and establish its formal connection to Fisher Information and optimal experimental design. We design a two-stage auto-bidding algorithm based on Lagrange duality that dynamically paces budget through a shadow price and optimizes impression-level bids using per-impression marginal utilities. To address missing labels at bid time, we propose a confidence-gated gradient heuristic, paired with a zeroth-order variant for black-box models that reliably estimates learning signals in real time. We provide theoretical guarantees, proving monotone submodularity of the composite objective, sublinear regret in online auction, and budget feasibility. Extensive offline experiments on synthetic and real-world datasets validate the framework: it outperforms baselines, achieves superior final AUC/LogLoss, adheres closely to budget targets, and remains effective when gradients are approximated zeroth-order. These results show that strategic, information-aware promotion can improve long-term model performance and organic outcomes beyond naive impression-maximization strategies.
Paper Structure (59 sections, 5 theorems, 64 equations, 10 figures, 1 algorithm)

This paper contains 59 sections, 5 theorems, 64 equations, 10 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Content Promotion Paradigm
  4. System Model.
  5. Problem Formulation.
  6. The Dilemma of Content Promotion
  7. An Empirical Performance Dichotomy in Content Promotion
  8. Redefining Impression Quality as Informative Data
  9. Optimal Experimental Design
  10. Methods
  11. Surrogate Uncertainty
  12. Advantages over classical metrics.
  13. Two-Stage Bidding
  14. Gradient Estimation
  15. Confidence-Gated Heuristic
  16. ...and 44 more sections

Key Result

Theorem 1

Let $\mathcal{D}_{\mathrm{val}}$ be a fixed validation set of size $k$, and for each $x\in\mathcal{D}_{\mathrm{val}}$ let $g(x)\in\mathbb{R}^d$ denote the loss gradient at a common anchor parameter $\theta_{\mathrm{anchor}}$. For a selected training set $S\subseteq\mathcal{D}_{\mathrm{train}}$ with and the regularized total uncertainty (analogy to lu2024daved) Let the gradient-coverage surrogate

Figures (10)

  • Figure 1: KPI Improvement ratio by Shutiao stratified by content CTR, compared with organic content.
  • Figure 2: System Diagram of Information-Aware Auto-Bidding.
  • Figure 3: Visualization of selected gradients.
  • Figure 4: Performance comparison of selected gradients on continually training. Our surrogate-based method achieves a significant reduction in Log Loss and an increase in AUC, closely approaching the performance of the FIM oracle and substantially outperforming the random baseline.
  • Figure 5: Evaluation of Budget Feasibility and Pacing Dynamics.
  • ...and 5 more figures

Theorems & Definitions (12)

  • Remark 1
  • Theorem 1: Regularized Fisher-Coverage Relationship
  • Theorem 2: Submorularity of the Uncertainty-Reduction Utility Function
  • Theorem 3: Regret Bound for First-Price CPM Dual Pacing
  • Theorem 4: Budget Feasibility Guarantee
  • Definition 1: Try-Accept Creator Strategy
  • Theorem 5: Non-Convex Convergence Rate with Noise
  • proof
  • proof
  • proof
  • ...and 2 more