Scalable and Traceable Joint Uplift Modeling for Multi-Lever Online Marketing

Abstract

In online marketing, platforms often adjust multiple incentive mechanisms simultaneously, such as discounts and coupons, to maximize user engagement and revenue. The interplay between these different categories of incentives is governed by joint effects, where their combined impact significantly deviates from a simple summation of individual effects. However, most existing uplift models focus primarily on single-category incentives; a straightforward extension of these models to multi-category scenarios leads to an explosion in parameter scale and a high risk of overfitting. To address this problem, this paper proposes a network named Multi-treatment Decomposable Low-Rank Interaction (M-DLRI). First, we design a joint monotonic encoding layer, which extends isotonic encoding to a joint space through linear interpolation and vector outer products. This allows the model to capture joint effects while ensuring strict joint monotonicity, aligning its output with economic common sense. Second, to handle more complex scenarios with a larger number of treatment categories, we derive a simple and elegant additive formula based on CP decomposition theory. This formula can efficiently process any treatment combination with high scalability and also supports causal traceability. Extensive evaluations on both synthetic datasets and a real large-scale industrial dataset confirm that M-DLRI outperforms state-of-the-art baseline models. Moreover, M-DLRI has been successfully deployed on the Alimama platform, delivering significant online gains, including a +5.35% lift in GMV and a +7.26% lift in Sales.