ADT4Coupons: An Innovative Framework for Sequential Coupon Distribution in E-commerce

Li Kong, Bingzhe Wang, Zhou Chen, Suhan Hu, Yuchao Ma, Qi Qi, Suoyuan Song, Bicheng Jin

Abstract

Coupon distribution is a critical marketing strategy used by online platforms to boost revenue and enhance user engagement. Regrettably, existing coupon distribution strategies fall far short of effectively leveraging the complex sequential interactions between platforms and users. This critical oversight, despite the abundance of e-commerce log data, has precipitated a performance plateau. In this paper, we focus on the scene that the platforms make sequential coupon distribution decision multiple times for various users, with each user interacting with the platform repeatedly. Based on this marketing scenario, we propose a novel marketing framework, named Aligned Decision Transformer for Coupons (ADT4Coupons), to directly devise coupon distribution policy for long-term revenue boosting. ADT4Coupons enables optimized online decision-making in a variety of real-world marketing scenarios. It achieves this by seamlessly integrating three key characteristics, general scenarios, sequential modeling with more comprehensive historical data, and efficient iterative updates within a unified framework. Furthermore, empirical results on real-world industrial dataset, alongside public and synthetic datasets demonstrate the superiority of our framework.

Introduction

In the competitive landscape of e-commerce, coupon distribution is a delicate balancing act: platforms must maximize long-term revenue and user engagement without exceeding finite marketing budgets. Traditionally, this has been approached through two-stage methods (uplift modeling followed by optimization) or single-round reinforcement learning. However, these methods often suffer from "error accumulation" or ignore the sequential nature of user behavior—where a coupon offered today fundamentally influences a user's likelihood to return tomorrow.

To bridge this gap, we introduce ADT4Coupons, implemented via the Sequence-Aware Constrained Optimization (SACO) framework. This approach treats coupon distribution as a sequential decision-making problem, adapting the Decision Transformer architecture to incorporate real-time budget constraints.

Methodology

The core innovation of ADT4Coupons is the Constrained-Optimization Decision Transformer. Unlike traditional models that view interactions in isolation, this framework models a user’s journey as a trajectory of states ($s$), actions ($a$), rewards ($r$), and costs ($c$).

The ADT4Coupons (SACO) architecture, featuring a Causal Transformer that processes trajectory sequences of states, actions, rewards, and costs while integrating budget constraints via dual variable embeddings.

Key Technical Components:

  • Trajectory Modeling: The model is trained on sequences of historical interactions, allowing it to capture temporal dynamics.
  • Constraint Integration: To handle strict budgets, the framework employs Lagrangian relaxation. A dual variable, $\lambda$, acts as a "price" for budget consumption and is fed directly into the Transformer as an embedding. This allows the model to adapt to different budget levels dynamically without retraining.
  • Conditioning on Goals: Following the Decision Transformer paradigm, the model is conditioned on "Return-to-Go" ($RTG$) and "Cost-to-Go" ($CTG$), enabling it to optimize for long-term revenue while staying within remaining funds.
  • Iterative Optimization: The framework utilizes the SLSQP (Sequential Least Squares Programming) algorithm to find the optimal $\lambda$ that maximizes revenue while ensuring the total cost remains under the budget limit.

Experimental Results

The framework was evaluated across four diverse datasets: CRITEO-UPLIFT v2 (CU), MT-LIFT (MT), a synthetic dataset (SY), and a large-scale industrial business dataset (BD) from ByteDance.

Performance and Efficiency

ADT4Coupons consistently outperformed state-of-the-art baselines, including Causal Forests, Bandits, and Decision Focused Causal Learning (DFCL). Key results include:

  • Revenue Boost: An average revenue improvement of 3.60% across all tested datasets.
  • Budget Efficiency: The model maintained a high Budget Allocation Rate (BARate), often reaching 99-100% efficiency without exceeding the cap.
  • Inference Speed: The system achieved an inference time of 3.8ms per user on an NVIDIA RTX 4090, meeting the low-latency requirements of real-time industrial production.
Revenue difference comparison on the MT dataset, demonstrating that the SACO framework consistently outperforms baselines across various budget levels.

Key Insights and Discussion

By embedding the dual variable $\lambda$ as a "control knob," the model learns a policy sensitive to budget pressure. When the budget is tight (high $\lambda$), the model becomes more selective, targeting only the most sensitive users. Furthermore, ADT4Coupons' auto-regressive structure reduces the heavy reliance on expensive Randomized Controlled Trial (RCT) data, allowing for updates using biased, logged historical data.

Future Directions

While ADT4Coupons represents a significant advancement, future work will address user-to-user externalities (such as social sharing) and explore methods to encode coupons so the system can adapt to entirely new discount types without full retraining. Additionally, research into more advanced distributed computing setups will help mitigate memory bottlenecks during massive-scale parallel user allocation.

Summary updated

Abstract

Coupon distribution is a critical marketing strategy used by online platforms to boost revenue and enhance user engagement. Regrettably, existing coupon distribution strategies fall far short of effectively leveraging the complex sequential interactions between platforms and users. This critical oversight, despite the abundance of e-commerce log data, has precipitated a performance plateau. In this paper, we focus on the scene that the platforms make sequential coupon distribution decision multiple times for various users, with each user interacting with the platform repeatedly. Based on this marketing scenario, we propose a novel marketing framework, named Aligned Decision Transformer for Coupons (ADT4Coupons), to directly devise coupon distribution policy for long-term revenue boosting. ADT4Coupons enables optimized online decision-making in a variety of real-world marketing scenarios. It achieves this by seamlessly integrating three key characteristics, general scenarios, sequential modeling with more comprehensive historical data, and efficient iterative updates within a unified framework. Furthermore, empirical results on real-world industrial dataset, alongside public and synthetic datasets demonstrate the superiority of our framework.
BibTeX
@inproceedings{kong2026adt,
  author = {Li Kong and Bingzhe Wang and Zhou Chen and Suhan Hu and Yuchao Ma and Qi Qi and Suoyuan Song and Bicheng Jin},
  title = {{ADT4Coupons: An Innovative Framework for Sequential Coupon Distribution in E-commerce}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i17.38525},
}
Figure