How Causal Inference Boosts Logistics Marketing Efficiency with Entropy Balancing

01 Causal Inference and Marketing Observational Data

Causal inference determines whether an event (cause) leads to another event (effect), distinguishing correlation from causation. In marketing, understanding how subsidies or pricing affect order growth is crucial. Observational data are abundant but suffer from confounding factors, requiring methods like propensity‑score matching, IPW, and entropy balancing.

Uplift models (Meta‑Learner, causal trees, deep learning) are used to estimate causal effects, demanding stability across scenarios and accurate effect learning to avoid negative returns.

02 Logistics Industry Technical Challenges

Key challenges include:

Fine‑grained control down to geographic fences.

Stability across diverse logistics scenarios.

Accurate uplift scores even when observational data quality degrades.

Time and space introduce imbalance: peak order periods, weekday/weekend differences, and spatial concentration in city centers versus suburbs, all affecting causal effect learning.

03 Spatio‑Temporal Entropy Balancing

A flexible grid method partitions time by order density and space using H3 hexagonal cells, aggregating cells to maintain sufficient order volume while preserving geographic continuity. Entropy balancing is applied within each spatio‑temporal cell, minimizing loss while controlling confounding.

Weights from entropy balancing are attached to observational data, then uplift models compute ITE and ATE, followed by optimization for marketing strategy deployment.

Gradient descent solves the weighted entropy balance, with PCA reducing confounding feature dimensionality and hyper‑parameters controlling grid granularity.

Experiments on real subsidy and pricing data show rapid loss convergence; AUC slightly drops while AUUC improves, indicating robust performance across datasets.

04 Summary and Outlook

The presented approach tackles practical challenges of using observational data for causal effect estimation in logistics marketing. The flexible grid entropy‑balancing method enhances confounding control and can be extended to other domains requiring fine‑grained causal inference.