Causal Inference with Observational Data for Improving Marketing Efficiency in the Logistics Industry

The presentation introduces a study on applying causal inference to observational marketing data in the logistics sector, aiming to improve the efficiency of marketing strategies such as subsidies, pricing, and user acquisition.

It explains that causal inference distinguishes true cause‑effect relationships from mere correlations, highlighting the importance of handling confounding factors, selection bias, and the need for stable, accurate uplift models across diverse scenarios.

The logistics industry faces technical challenges including fine‑grained geographic targeting, model stability across varied operational contexts, and maintaining high uplift score accuracy despite noisy observational data.

To address these, the authors propose a spatiotemporal entropy‑balancing method that partitions data using flexible grids based on order density, employing H3 hexagonal cells aggregated into adaptable regions to ensure sufficient sample size while preserving spatial continuity.

Entropy balancing is extended to each spatiotemporal cell, minimizing information loss and controlling confounding bias; weights are optimized via gradient descent with PCA‑based dimensionality reduction and regularization to prevent gradient explosion.

The method integrates the computed weights into uplift models to estimate individual and average treatment effects, which are then combined with operational optimization to allocate marketing interventions effectively.

Experimental results on real subsidy and pricing data, as well as simulated datasets, show rapid loss convergence, improved causal effect estimation, and better handling of confounding compared to traditional approaches, with favorable AUC/AUUC metrics.

In conclusion, the flexible grid entropy‑balancing framework provides a scalable solution for precise causal effect learning in logistics marketing and can be adapted to other domains requiring fine‑grained confounding control.