Causal Inference and Entropy Balancing for Improving Marketing Efficiency in the Logistics Industry

The article introduces a practical application of causal inference to improve marketing efficiency in the logistics sector, using observational data instead of costly randomized experiments. It explains why causal reasoning is essential for marketing decisions such as subsidies and pricing, and outlines the challenges of confounding, selection bias, and OOD generalization.

It describes the logistics industry's technical challenges, including the need for fine‑grained, stable, and high‑performing models that can handle diverse geographic fences, time‑varying demand patterns, and varying order densities across regions and time slots.

To address these challenges, the authors propose a spatio‑temporal entropy‑balancing method built on flexible grid partitioning. The approach first divides the data into time slices based on order density and then creates adaptive spatial grids using the H3 hexagonal indexing system, aggregating cells to ensure sufficient order volume while preserving geographic continuity.

Within each spatio‑temporal cell, entropy‑balancing weights are computed to minimize confounding influence, constrained by smoothness and balance requirements. These weights are applied to the observational data, after which an uplift model (meta‑learner, causal tree, or deep learning variant) estimates individual and average treatment effects, which are then integrated with operational optimization to allocate marketing interventions.

The implementation includes techniques such as inverse propensity weighting (IPW), entropy balancing, PCA for dimensionality reduction, and gradient‑based optimization with Lagrangian dual methods to solve the weighted loss. The method also incorporates business constraints like monotonicity of subsidies and pricing.

Experimental validation using real subsidy and pricing data, as well as simulated datasets, shows rapid loss convergence, improved AUC/AUUC metrics, and better confounding control compared to traditional methods. The spatio‑temporal entropy‑balancing approach yields higher uplift scores, lower p‑values, and more reliable causal effect estimates across fine‑grained grids.

In conclusion, the presented framework demonstrates how observational data, when combined with causal inference and flexible grid‑based entropy balancing, can effectively enhance marketing strategies in logistics and is extensible to other domains requiring fine‑grained causal analysis.