[Paper Review] An End-to-End Framework for Marketing Effectiveness Optimization under Budget Constraint

Ziang Yan, Shusen Wang+3 more|arXiv (Cornell University)|Feb 9, 2023

Consumer Market Behavior and Pricing4 citations

TL;DR

This paper proposes an end-to-end deep learning framework that directly optimizes marketing effectiveness under budget constraints by integrating a differentiable regularizer based on the Expected Outcome Metric (EOM) for the multiple choice knapsack problem. By using gradient estimation techniques like finite-difference and natural evolution strategies (NES), the model jointly optimizes treatment effect prediction and business goal maximization, achieving significant improvements in both offline and online A/B tests on a large-scale short video platform.

ABSTRACT

Online platforms often incentivize consumers to improve user engagement and platform revenue. Since different consumers might respond differently to incentives, individual-level budget allocation is an essential task in marketing campaigns. Recent advances in this field often address the budget allocation problem using a two-stage paradigm: the first stage estimates the individual-level treatment effects using causal inference algorithms, and the second stage invokes integer programming techniques to find the optimal budget allocation solution. Since the objectives of these two stages might not be perfectly aligned, such a two-stage paradigm could hurt the overall marketing effectiveness. In this paper, we propose a novel end-to-end framework to directly optimize the business goal under budget constraints. Our core idea is to construct a regularizer to represent the marketing goal and optimize it efficiently using gradient estimation techniques. As such, the obtained models can learn to maximize the marketing goal directly and precisely. We extensively evaluate our proposed method in both offline and online experiments, and experimental results demonstrate that our method outperforms current state-of-the-art methods. Our proposed method is currently deployed to allocate marketing budgets for hundreds of millions of users on a short video platform and achieves significant business goal improvements. Our code will be publicly available.

Motivation & Objective

• To address the misalignment between treatment effect estimation and budget allocation objectives in two-stage marketing optimization pipelines.
• To directly optimize the business goal (e.g., user response rate) under budget constraints in a unified end-to-end framework.
• To enable deep neural networks to learn from the actual marketing outcome metric rather than relying on proxy objectives in treatment effect estimation.
• To overcome the non-differentiability of integer programming-based objective functions in budget allocation via gradient estimation.
• To deploy and validate the method at scale on a real-world short video platform with hundreds of millions of users.

Proposed method

• Formulates the budget allocation problem as a multiple choice knapsack problem (MCKP) with user-level treatment effects and costs.
• Constructs an unbiased estimator $Q$ of the expected business outcome using the Expected Outcome Metric (EOM) method.
• Integrates $Q$ as a differentiable regularizer into the S-Learner neural network training objective to directly optimize for marketing effectiveness.
• Employs finite-difference and natural evolution strategies (NES) to estimate gradients of the non-differentiable $Q$ function for backpropagation.
• Uses a hyperparameter $\lambda$ to balance the predictive loss and the marketing goal regularizer in the joint optimization.
• Applies the framework to real-world marketing campaigns, with model retraining and deployment on a large-scale short video platform.

Experimental results

Research questions

• RQ1 Can an end-to-end deep learning framework outperform two-stage approaches in marketing budget allocation by directly optimizing the business goal?
• RQ2 How can the non-differentiable nature of integer programming-based objectives be addressed to enable gradient-based optimization in marketing effectiveness?
• RQ3 What is the impact of different gradient estimation strategies (e.g., finite-difference vs. NES) on optimization performance and computational cost?
• RQ4 How does the hyperparameter $\lambda$ controlling the regularizer influence model stability and performance in real-world deployment?
• RQ5 Can the proposed method achieve measurable improvements in both offline metrics and online A/B tests on a large-scale platform?

Key findings

• On the CRITEO-UPLIFT v2 dataset, the proposed method with finite-difference gradient estimation achieved an AUCC of 0.7904 ± 0.0096, significantly outperforming the previous state-of-the-art method (DRP) with 0.7739 ± 0.0002.
• The method with finite-difference gradient estimation required 17.6 hours on an NVIDIA Tesla T4 GPU, while the NES-based variant completed in 4.3 hours, indicating a 4x speedup trade-off with higher variance.
• In offline evaluation on the KUAISHOU-PRODUCE-COIN dataset, the optimal $\lambda = 200$ yielded the highest per-capita response, while $\lambda \geq 500$ led to training instability and worse performance than TSM-SL.
• In online A/B testing, the proposed method achieved a 1.24% higher per-capita response and a 0.369% lower per-capita cost compared to TSM-SL, demonstrating significant business impact.
• The model is currently deployed to allocate marketing budgets for hundreds of millions of users on a short video platform, confirming scalability and real-world effectiveness.
• The results show that directly optimizing the business goal via a differentiable regularizer leads to better alignment with real-world objectives than traditional two-stage pipelines.