Agent-Based Simulation and Eleven-Stage ML Pipeline Code for Climate-Adaptive Practice Adoption Among Smallholder Farmers (Zone K)

Abstract

Quantifying Social Diffusion in Climate-Adaptive Practice Adoption: An Agent-Based Simulation and Explainable Machine Learning Framework for Smallholder Farmers." It contains the agent-based simulation generator and the complete eleven-stage analysis pipeline used to produce every table, figure, and reported statistic in the article: farmer segmentation (Gaussian Mixture Model), discrete-time hazard and ensemble classification of practice adoption, mixed-effects and gradient-boosted yield regression, spatial-lag logistic and GraphSAGE graph-neural-network models of network effects, the Social Diffusion Attribution Score (SHAP-based), feature-group ablation, naive-baseline comparison, paired statistical significance testing, and a dataset-size robustness check, along with the scripts used to generate the manuscript's figures. The companion synthetic dataset (the simulated farmer, network, panel, and narrative data itself) is deposited separately at https://zenodo.org/records/22785299. Running the scripts in this archive against that dataset, or against a freshly regenerated copy produced by the included generator script, reproduces the study's results in full. No real farmer, household, or personally identifiable data is used or represented anywhere in this study; both archives are entirely synthetic/simulation-derived.