How Combining Causal Inference with Genetic Algorithms Optimizes Freight Pricing

Introduction

In the global economy, freight services are crucial, and effective pricing strategies are key to competitiveness. Traditional integer programming and causal inference struggle with data limitations and model bias, especially in the dynamic e‑commerce environment.

Industry Solution Overview

Recent industry work combines causal inference (CI) with integer programming (IP) to estimate heterogeneous treatment effects of price factors and solve optimization problems under budget constraints. However, these methods suffer from exponential runtime growth and sensitivity to prediction errors.

Proposed CI‑GA Framework

The authors propose integrating causal inference with a genetic algorithm (GA) to reduce dependence on data quality. GA provides global search capability without strict data assumptions, while CI offers precise effect estimation, together yielding robust solutions even with biased or limited data.

Causal Inference Model Application

Data from AB experiments on price policies are used, with features grouped into four categories: order‑level, real‑time scene, user‑level, and price‑policy features. Meta‑learners such as T‑learner are employed to estimate uplift, as tree‑based methods suit the discrete, low‑dimensional data.

Genetic Algorithm Application

GA simulates natural selection, using crossover, mutation, and selection to evolve populations of candidate solutions. It has been successfully applied to hyper‑parameter optimization (e.g., DeepMind’s PBT) and feature selection, demonstrating superior performance in large‑scale online markets.

CI‑GA Framework Details

The framework consists of model construction (non‑linear and hybrid causal models), algorithm design (causal priors, deep learning integration), search strategy (global GA search with causal validity assessment), and feedback mechanisms (fitness functions incorporating CI results and online iterative updates).

Offline Simulation

Experiments on synthetic datasets compare CI‑GA with CI‑IP across four data conditions (large/small samples, biased/unbiased). Results show CI‑GA achieves better constraint satisfaction and pairing performance, especially under biased or small‑sample scenarios.

Case Study: Simulated Driver Commission Reduction

The study models commission discounts as treatments affecting driver‑passenger matching rates. Using simulated features (order, scene, user, price policy), the CI‑GA approach optimizes discount allocation under budget limits, outperforming IP in robustness and ROI.

Conclusion

CI‑GA outperforms CI‑IP in non‑ideal data conditions, offering robustness, global search ability, dynamic adjustment, causal validity, and adaptability, making it well‑suited for real‑world freight pricing optimization.

References

Wan et al., "Generalized Causal Forest..." (2022)

Ai et al., "Large‑Scale Budget‑Constrained Causal Forest" (WWW'22)

Liu et al., "Explicit Feature Interaction‑aware Uplift Network" (2023)

Chen et al., "Two‑Sided Instant Incentive Optimization" (ICDE 2023)

Ben‑Tal et al., "Robust Optimization" (2009)

Shi et al., "Adapting Neural Networks for Treatment Effects" (NeurIPS 2019)

Nie et al., "Vcnet and Functional Targeted Regularization" (2021)

Wang et al., "Secure Your Ride" (IEEE TKDE 2021)

Jaderberg et al., "Population Based Training" (2017)