Computational Model for Homeowner Choice Optimization

Computational Model for Homeowner Choice Optimization Authors: Tanjeel Ahmed Bin Zaman; Tabassum Haque; Himadri Sen Gupta; Mojtaba Harati; Niamat Ullah Ibne Hossain; John W. van de Lindt This deposit provides materials to reproduce the results of the manuscript "A predict-then-select framework for flood mitigation program menus under household participation uncertainty." The paper develops a predict-then-select pipeline that (i) predicts homeowner mitigation uptake using a leakage-safe CatBoost classifier with post-hoc Dirichlet calibration, and (ii) embeds the calibrated probabilities in an enumeration-based menu-selection framework to choose cost-effective and (optionally) equity-aware flood mitigation program menus. Provenance and scope Source data: The survey data used in this study are publicly available via DesignSafe-CI (DOI: 10.17603/ds2-ebpq-6s93). Study context: Stated-preference discrete choice experiment with 1,366 U.S. coastal property owners across three waves (Wave 1: Aug–Sep 2018; Wave 2: Oct–Nov 2018; Wave 3: Mar–Apr 2019), comprising 6,234 choice tasks. Behavioral model: Gradient-boosted decision trees (CatBoost) with Dirichlet calibration to improve probability quality for downstream decision-making. Outcome-derived choice variables are removed to prevent leakage. Optimization: Enumeration-based selection over 30 candidate program menus, maximizing expected mitigation uptake under budget constraints, with optional equity constraints based on income-group uptake. Key reported outcomes: Because the binary outcome is highly imbalanced (80.8% of choices involve some mitigation), raw accuracy is not the primary criterion. After removing outcome-derived variables, the Dirichlet-calibrated CatBoost model reaches 80.7% accuracy and ≈0.444 log loss, close to the 80.8% majority-class baseline; its value lies in improved probability quality (lower log loss and expected calibration error) rather than raw accuracy. The optimized uptake–budget relationship exhibits a stepwise frontier (expected uptake rising from 70.0% to 81.8% across the evaluated budget range) with diminishing returns at higher budgets. Contents replication_notebook.ipynb — a single, self-contained notebook that reproduces the full analysis end to end: feature engineering and grouped (respondent-aware) cross-validation; CatBoost training with Dirichlet calibration and probability-quality diagnostics (accuracy, balanced accuracy, macro-F1, log loss, ECE); and menu evaluation, the budget–uptake frontier, equity-constrained runs, robustness checks, and bootstrap uncertainty. The notebook installs any missing packages on first run. data/ ECUCNHR_Finaldata_05-17-2021_preprocessed.csv — analysis-ready preprocessed data used by the notebook (required). ECUCNHR_Finaldata_05-17-2021.csv — raw survey extract (optional). Raw data can also be downloaded from the DesignSafe DOI above. outputs/ — exported tables, figures, and fitted models generated by the notebook (choice/menu distributions, cross-validated metrics, frontier tables, calibration and SHAP diagnostics, equity and robustness results), archived as replication_outputs.zip. Reproducibility notes All random operations use a fixed seed (RANDOM_STATE = 42). Validation is respondent-grouped throughout, so no respondent appears in both training and test folds. A CPU runtime is sufficient (no GPU required). Minor variation (≤0.01–0.03) may appear in the CatBoost classification metrics across library versions and platforms; the budget frontier and menu selections are reproduced exactly.