Multi-Dimensional Causal Forest Model for Heterogeneous Treatment Effects in Marketing

This research presents a new HTE estimation method called multi-dimensional causal forest model, integrating efficient integer programming algorithms to address marketing decision-making challenges. The approach demonstrates superior performance over existing tree-based models by simultaneously handling multiple interventions and optimizing resource allocation under constraints.

The study tackles the core problem of heterogeneous treatment effects, where the same marketing intervention yields varying results across different user groups. By combining causal forest structures with parallelized dual gradient bisection algorithms, the method achieves efficient computation for billion-scale user bases while maintaining solution quality.

Key innovations include a two-step splitting algorithm balancing inter-node and intra-node heterogeneity, parallelizable integer programming solutions for large-scale optimization, and a novel RCT-based evaluation framework for offline model assessment. The approach has been validated through both offline simulations and real-world A/B experiments in Tencent's marketing operations, showing significant performance improvements.

Code implementations and supplementary materials are available at

https://github.com/www2022paper/WWW-2022-PAPER-SUPPLEMENTARY-MATERIALS

and the paper is published at https://arxiv.org/abs/2201.12585.