Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds

Chenfei Li, Hantao Zhao, Weixi Yao, Ruiming Huang, Rongrong Lu, Geng Tian, Dongying Kong

TL;DR

This work proposes a constraint-aware generative reranking framework that transforms constrained optimization into bounded neural decoding, and introduces constraint-aware reward pruning, integrating constraint satisfaction directly into decoding to efficiently generate optimal sequences.

Abstract

Optimizing reranking in advertising feeds is a constrained combinatorial problem, requiring simultaneous maximization of platform revenue and preservation of user experience. Recent generative ranking methods enable listwise optimization via autoregressive decoding, but their deployment is hindered by high inference latency and limited constraint handling. We propose a constraint-aware generative reranking framework that transforms constrained optimization into bounded neural decoding. Unlike prior approaches that separate generator and evaluator models, our framework unifies sequence generation and reward estimation into a single network. We further introduce constraint-aware reward pruning, integrating constraint satisfaction directly into decoding to efficiently generate optimal sequences. Experiments on large-scale industrial feeds and online A/B tests show that our method improves revenue and user engagement while meeting strict latency requirements, providing an efficient neural solution for constrained listwise optimization.

Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds

TL;DR

This work proposes a constraint-aware generative reranking framework that transforms constrained optimization into bounded neural decoding, and introduces constraint-aware reward pruning, integrating constraint satisfaction directly into decoding to efficiently generate optimal sequences.

Abstract

Optimizing reranking in advertising feeds is a constrained combinatorial problem, requiring simultaneous maximization of platform revenue and preservation of user experience. Recent generative ranking methods enable listwise optimization via autoregressive decoding, but their deployment is hindered by high inference latency and limited constraint handling. We propose a constraint-aware generative reranking framework that transforms constrained optimization into bounded neural decoding. Unlike prior approaches that separate generator and evaluator models, our framework unifies sequence generation and reward estimation into a single network. We further introduce constraint-aware reward pruning, integrating constraint satisfaction directly into decoding to efficiently generate optimal sequences. Experiments on large-scale industrial feeds and online A/B tests show that our method improves revenue and user engagement while meeting strict latency requirements, providing an efficient neural solution for constrained listwise optimization.
Paper Structure (50 sections, 1 theorem, 33 equations, 2 figures, 3 tables)

This paper contains 50 sections, 1 theorem, 33 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reranking and Listwise Optimization
  4. Generative Recommendation and Autoregressive Ranking
  5. Constrained Allocation in Advertising Systems
  6. Our Contribution
  7. Problem Formulation
  8. Candidate Set
  9. Ranking Sequence
  10. List-Level Reward
  11. Business Constraints
  12. (1) Advertisement Load Constraint
  13. (2) Position Spacing Constraint
  14. (3) Structural Rules
  15. Constrained Optimization Objective
  16. ...and 35 more sections

Key Result

Theorem 9.1

Let $\mathcal{F}$ denote the feasible set of ranking sequences satisfying all business constraints, including advertisement load constraints and positional spacing constraints. Assume that the maximum number of advertisements per list is bounded by $K$. If the bounded decoding procedure enumerates a

Figures (2)

  • Figure 1: Model struct of CGR.
  • Figure 2: Two-stage constraint-aware generative inference framework. The first stage performs constrained insertion over the natural content list and selects a top-1 candidate using the reward model. The second stage performs bounded generative decoding with large-ad candidates, enumerating feasible sequences (single-ad, double-ad, and no-ad lists) under constraint rules. Constraint-aware reward pruning removes infeasible or suboptimal sequences, and the final list is selected via reward maximization.

Theorems & Definitions (2)

  • Theorem 9.1: Optimality of Bounded Decoding
  • proof