This repository provides an explainable machine learning approach to analyze factors affecting fertility rates across various regions in Korea. Using SHAP (SHapley Additive exPlanations) values, we identify the most critical features and their contributions to model predictions.
The project uses two state-of-the-art gradient boosting algorithms: XGBoost and LightGBM, to predict fertility rates. These predictions are explained with SHAP, which allows us to understand the impact of different socio-economic and regional factors.
We used open public data from KOSIS, ranging Social/Economical/Cultural/Regional/and other factors that describe Korean society from 2015-2023. Some data include:
| Region | Year | GRDP per Capita (2015-2022) | Regular Workdays (Days) | Regular Work Hours (Hours) | Monthly Salary (KRW) | Fixed Salary (KRW) | Overtime Hours (Hours) | Overtime Salary (KRW) | Total Work Hours (Hours) | Special Bonus (KRW) | Total Fertility Rate (2015-2023) | Single-Person Household Ratio | Health Practice Rate (2015-2023) | Economic Participation Rate (2015-2023) | Employment Rate (2015-2023) | High-Risk Drinking Rate (2015-2023) | Aging Index (2015-2023) | Cultural Facilities per 100K People (2015-2022) | Cultural Activity Population Ratio (2015-2023) | Private Education Expenses | Gender Equality Index (2017-2022) | Income (2015-2022) | Stress Awareness Rate (2015-2023) | Unemployment Rate (2015-2023) | Female Economic Participation Rate (2015-2023) | Female Employment Rate (2015-2023) | Female Unemployment Rate (2015-2023) | Female Youth Ratio (19-39) (2015-2023) | Female Average Marriage Age (2015-2023) | Average Monthly Wage (2015-2023) | Drinking Rate (2015-2023) | Population Density (2015-2023) | Crude Divorce Rate (2015-2023) | Crude Marriage Rate | Youth Ratio (19-39) (2015-2023) | Average Age (2015-2023) |
|---------|------|-----------------------------|--------------------------|-----------------------------|-----------------------|--------------------|--------------------------|------------------------|----------------------------|---------------------|---------------------------------|-------------------------------|----------------------------------|--------------------------------|------------------------|--------------------------------|-------------------------|------------------------------------------|-----------------------------------------|------------------------|----------------------------------|----------------------|--------------------------------|----------------------------------|------------------------|--------------------------------|----------------------------------|-------------------------|--------------------------|----------------------------|--------------------------|-----------------------|-------------------------|-------------------|
| Gangwon | 2015 | 27608 | 22.3 | 174.0 | 2569139 | 2569139 | 162.7 | 1627.8 | 192.4 | 140000 | 1.18 | 19.8 | 25.5 | 53.3 | 45.4 | 7.3 | 15.2 | 8.4 | 13.2 | 124.7 | 0.5 | 123 | 4.3 | 32.8 | 45.3 | 50.5 | 29.3 | 2.1 | 63.5 | 182.0 | 3.1 | 19.1 |
| Gangwon | 2016 | 29380 | 21.0 | 170.0 | 2700000 | 2700000 | 157.3 | 1400.0 | 184.6 | 142000 | 1.20 | 21.2 | 25.8 | 54.1 | 46.0 | 6.8 | 16.1 | 9.2 | 14.1 | 133.3 | 0.6 | 130 | 4.0 | 34.2 | 46.1 | 52.0 | 30.0 | 2.0 | 62.8 | 190.0 | 2.9 | 19.3 |
| Gangwon | 2017 | 31200 | 20.8 | 165.3 | 2830000 | 2830000 | 151.2 | 1300.0 | 176.5 | 150000 | 1.23 | 22.5 | 26.2 | 55.0 | 47.5 | 6.5 | 17.0 | 10.1 | 15.3 | 140.8 | 0.7 | 140 | 3.8 | 35.5 | 47.3 | 53.2 | 30.8 | 1.9 | 61.7 | 198.0 | 2.7 | 19.6 |
| Gangwon | 2018 | 32267 | 21.3 | 169.4 | 2980000 | 2980000 | 153.0 | 1500.0 | 180.4 | 160000 | 1.26 | 23.1 | 26.6 | 56.0 | 48.0 | 6.1 | 18.1 | 10.9 | 16.3 | 148.2 | 0.8 | 145 | 3.5 | 36.8 | 48.0 | 54.2 | 31.5 | 1.8 | 60.3 | 206.0 | 2.6 | 19.8 |
The SHAP feature importance plots illustrate the overall impact of features:
-
LightGBM SHAP Importance
-
XGBoost SHAP Importance
These plots show how specific features influence the model's prediction for fertility rates:
-
Average Age at First Marriage
-
Average Age
-
Average Monthly Salary
-
Crude Divorce Rate
-
Drinking Rate
explainable_fertility-main/
├── data/ # Raw and preprocessed data files
├── output_metrics/ # Evaluation metrics for models
├── result/ # Plots and visualizations
├── shap_data_preprocess.ipynb # Preprocessing code for SHAP
├── shap_model.ipynb # SHAP model training and analysis
└── shap_legacy.ipynb # Legacy SHAP experiments
This project is licensed under the MIT License.