Towards Effective and Efficient Multi-valued Treatment Uplift Modeling in Online Marketing

The effectiveness of an online marketing campaign heavily relies on the identification of user groups that exhibit high sensitivity to specific treatments. However, existing works in this domain has encountered certain limitations when applied in practical settings. Firstly, most studies have primarily focused on binary treatment scenarios, but real-world industrial applications often involve multi-valued treatments, rendering these approaches incompatible. Secondly, although a few studies have addressed multi-valued treatment scenarios, many of them have directly extended binary treatment architectures without considering additional optimization. This oversight can result in redundant model parameters and performance bottlenecks. In order to encounter aforementioned challenges, we propose a novel reparameterization multi-head treatment uplift network, or RMNet for short. RMNet incorporates an invariant feature representation and a reparameterization multi-head module. This module achieves a balanced representation of all treatments by employing gradient constraints, thereby mitigating selection bias and enhancing model efficiency and performance. The latter responses to different treatments as offsets relative to the control response, thus we employ a reparameterization multi-head structure to effectively reduce the number of model parameters necessary for predicting responses to different treatments. Finally, extensive experiments are conducted on two datasets to demonstrate the effectiveness and efficiency of our RMNet.