🌍 Environmental Livability Analysis — Country-Level Assessment (2019)

Author: Junaid Ahmed Rupok
Venue: IEEE Journal (Methodology Paper)
Dataset: 175 countries · 24 attributes · Year 2019

---

📌 Overview

This project presents an end-to-end machine learning pipeline for assessing environmental livability at the country level. The framework integrates unsupervised structure discovery, supervised classification, statistically robust feature attribution, and policy-oriented counterfactual generation — all grounded in a rigorous mathematical foundation formalized in an IEEE-format methodology paper.

The pipeline answers three core questions:

1. Which countries are environmentally livable, and why?
2. Which environmental features are statistically significant predictors of livability?
3. What minimal policy interventions can transition a non-livable country toward livability?

---

🗂️ Repository Structure

├── Environment_Project.ipynb          # Full implementation (30 cells)
├── country_environmental_livability_2019.csv  # Dataset (175 countries, 24 features)
├── methodology.tex                    # IEEE LaTeX methodology paper
└── README.md

---

🔬 Pipeline Architecture

The analytical framework comprises six sequential phases:

Raw Data (N=175, p=24)
        │
        ▼
┌─────────────────────────────┐
│  Phase 1: Preprocessing     │  Quality checks · IQR outlier detection
│                             │  Adaptive transforms · VIF removal · Standardization
└────────────┬────────────────┘
             ▼
┌─────────────────────────────┐
│  Phase 2: Unsupervised      │  PCA (≥85% variance) · K-Means clustering
│  Structure Discovery        │  Stability via ARI · EQ scoring · Livability labeling
└────────────┬────────────────┘
             ▼
┌─────────────────────────────┐
│  Phase 3: Supervised        │  5-model benchmarking · XGBoost selection
│  Classification             │  Grid search hyperparameter tuning · Stratified 5-fold CV
└────────────┬────────────────┘
             ▼
┌─────────────────────────────┐
│  Phase 4: ABR-SHAP          │  Bootstrap-retrained SHAP · Convergence testing
│  Feature Importance         │  95% CIs · Cohen's d · Overstatement Ratio
└────────────┬────────────────┘
             ▼
┌─────────────────────────────┐
│  Phase 5: CMCS Validation   │  Cross-methodological consistency
│                             │  Permutation testing (N=5000) · p-value · Effect size
└────────────┬────────────────┘
             ▼
┌─────────────────────────────┐
│  Phase 6: Pareto            │  Statistically-constrained counterfactuals
│  Counterfactuals            │  Minimum-cost interventions · Policy feasibility
└─────────────────────────────┘

---

📊 Figures

All 18 figures generated by the pipeline are presented below. The 10 most important figures are displayed inline; the remaining figures are referenced by name and description.

---

Fig. 1 — Outlier Analysis

fig_outlier_analysis.png

IQR-based outlier detection across all 24 features. Countries flagged with ≥5 outlier attributes (ARE, FIN, ISL, ISR, KWT) are identified for sensitivity analysis. gdp_per_capita (30 outliers) and life_expectancy (19 outliers) are the most affected features.

---

Fig. 2 — PCA Analysis

fig_pca_analysis.png

Scree plot, cumulative explained variance, PC1/PC2 loadings heatmap, and biplot. 11 components explain 87.7% of total variance (exceeding the 85% threshold). PC1 (28.5%) captures the human development–energy axis; PC2 (13.3%) reflects land use and ecological patterns.

---

Fig. 3 — K-Means Clustering Validation

fig_clustering_validation.png

Multi-metric cluster validation (Elbow, Silhouette, Davies-Bouldin, Composite Score) over K ∈ {2, …, 10} with 10 runs per K. All four metrics unanimously select K = 2. Silhouette peaks at 0.286 and Davies-Bouldin reaches its minimum at 1.534 for K = 2.*

---

Fig. 4 — Final Clustering & Stability Assessment

fig_final_clustering.png

Two-cluster partition in PCA space and ARI stability distribution across 100 independent K-Means runs. Mean ARI = 1.000 (std = 0.000) confirms perfectly deterministic and replicable partitioning. Cluster 0: 42 countries (24%); Cluster 1: 133 countries (76%).

---

Fig. 5 — Environmental Quality Scoring & Livability Labeling

fig_env_quality_livability.png

Composite Environmental Quality (EQ) scores by cluster and resulting livability label assignment. Cluster 1 (mean EQ = 0.633) is designated livable (y = 1); Cluster 0 (mean EQ = 0.549) is non-livable (y = 0). Final distribution: 133 livable (76%), 42 non-livable (24%).

---

Fig. 6 — Model Benchmarking

fig_model_benchmarking.png

Stratified 5-fold cross-validation comparison of five classifiers across Accuracy, F1, Precision, Recall, and AUC-ROC. Logistic Regression achieves the highest CV F1 (0.991). XGBoost (CV F1 = 0.973, AUC = 0.982) is selected for final modeling due to native TreeSHAP compatibility.

---

Fig. 7 — ABR-SHAP Feature Importance

fig_abr_shap_results.png

Bootstrap-retrained SHAP importance rankings with 95% confidence intervals (B = 200 iterations) and Cohen's d effect sizes. 9 of 19 features are statistically significant (lower CI > 0, shown in orange). Overstatement Ratio = 2.11, meaning standard SHAP inflates the significant feature set by a factor of ~2. life_expectancy leads with mean |SHAP| ≈ 0.84 and Cohen's d ≈ 2.7.*

---

Fig. 8 — SHAP Summary & Dependence Plots

fig_shap_detailed.png

Standard SHAP bar chart, beeswarm summary plot, and dependence plots for life_expectancy and co2_per_capita. High life_expectancy values (red, beeswarm right) strongly increase livability probability. co2_per_capita shows non-linear suppression of livability at high emission levels.

---

Fig. 9 — CMCS Validation

fig_cmcs_validation.png

Cross-Methodological Consistency Score vs. null permutation distribution (N = 5,000). Observed CMCS = 0.400 lies far beyond the 95th percentile of the null (0.058), with p = 0.0002 and Cohen's d = 14.43. All 9 significant features individually achieve per-feature silhouette scores with p = 0.0002, confirming strong supervised–unsupervised structural alignment.

---

Fig. 10 — Cluster Profiles Heatmap

cluster_profiles_heatmap.png

Full normalized feature profiles for both clusters. Cluster 0 (non-livable) scores higher on CO₂ per capita, fossil fuel dependency, electricity access, urban population, and economic activity. Cluster 1 (livable) shows higher forest area, energy intensity, and trade openness — reflecting a lower-industrialization, higher-ecological-integrity profile.*

---

Fig. 11 — VIF Removal Convergence

fig_vif_removal.png

VIF convergence trajectory across iterations and final retained feature VIF values. renewable_energy_pct (VIF = 30.6) and gdp_per_capita (VIF = 11.4) are removed in two iterations. All 19 retained features satisfy VIF ≤ 10 (max: co2_per_capita at 7.09).

---

Fig. 12 — Adaptive Transformation Effects

fig_transformation_effects.png

Before/after skewness distributions for the 5 adaptively transformed features. Log transformations reduce co2_per_capita skewness from 2.97 → 0.57, carbon_damage from 2.61 → 1.25, gdp_per_capita from 2.16 → 0.24, life_expectancy from 1.69 → 1.45, and air_transport from 1.42 → −0.02.

---

Fig. 13 — Feature Pre-Screening

fig_feature_prescreening.png

Pearson |r| and Mutual Information rankings of all features with respect to CO₂ per capita. carbon_damage leads both metrics (|r| = 0.857, MI = 0.698). Features like fertilizer_consumption, air_transport, and manufacturing_pct show near-zero MI, foreshadowing their exclusion by ABR-SHAP.

---

Fig. 14 — External Validation

fig_external_validation.png

Pearson and Spearman correlations between predicted livability probabilities and CO₂ per capita as an external proxy (r = 0.033, p = 0.852), with 95% bootstrap CI [−0.295, 0.209]. The near-zero correlation reflects that livability is a multi-dimensional construct not reducible to a single emission indicator.

---

Fig. 15 — Partial Correlation Analysis

fig_partial_correlation.png

Partial correlations between ABR-SHAP significant features and livability, controlling for GDP per capita. For most features, partial r closely tracks original r, confirming that feature–livability associations are not confounded by economic development alone.

---

Fig. 16 — Temporal Stability / Noise Sensitivity

fig_stability_analysis.png

Feature overlap with the original ABR-SHAP significant set across five Gaussian noise levels (σ ∈ {0.05, …, 0.25}), each with 20 trials. Overlap remains consistently above 82% with no monotonic decay, confirming robustness to measurement noise.

---

Fig. 17 — Feature Group Ablation Study

fig_ablation_study.png

Stratified 5-fold F1 scores when the model is restricted to each feature domain. Governance features (3 features, F1 = 0.931) achieve the highest efficiency at 0.311 F1/feature. Environmental (0.936) and Social (0.945) features are competitive. Economic features trail at F1 = 0.828.

---

Fig. 18 — Counterfactual & Policy Analysis

fig_counterfactuals.png · policy_recommendations.png

Per-country Pareto counterfactual perturbations for 8 non-livable test countries. Only 1 country (TJK — Tajikistan) achieves a feasible single-feature intervention: a carbon_damage reduction of −1.9σ yields livability probability 0.536. The 87.5% infeasibility rate reflects the structural depth of environmental challenges in the most stressed nations.

---

📐 Novel Methodological Contributions

1 · Adaptive Bootstrap-Retrained SHAP (ABR-SHAP)

Standard SHAP values are computed from a single model fit and can overstate feature importance by conflating signal with sampling variability. ABR-SHAP addresses this by:

• Retraining the model on B bootstrap samples (adaptive convergence, B_max = 500)
• Computing SHAP values on each retrained model
• Reporting mean importance ± 95% bootstrap CI and Cohen's d per feature
• Flagging features significant only when the lower CI bound exceeds zero
• Quantifying the Overstatement Ratio (OR) between standard and bootstrap-significant features

2 · Cross-Methodological Consistency Score (CMCS)

CMCS validates that the features identified by supervised SHAP analysis independently explain the structure discovered by unsupervised clustering:

$$\text{CMCS}_{\text{obs}} = \frac{1}{|\mathcal{F}_{\text{sig}}|} \sum_{f \in \mathcal{F}_{\text{sig}}} \text{sil}(f, \mathbf{c})$$

Statistical significance is established via 5,000 permutation tests with a bias-corrected p-value.

3 · Statistically-Constrained Pareto Counterfactuals

For each non-livable country, the algorithm finds the minimum-cost single-feature perturbation that shifts the XGBoost prediction past the decision boundary — but only over the statistically significant feature set $\mathcal{F}_{\text{sig}}$. Interventions are bounded to ±3σ and flagged as feasible if |δ*| ≤ 2.4σ.

---

📈 Key Results Summary

Metric	Value
Dataset	175 countries, 24 features, Year 2019
PCA components (≥85% variance)	11 (87.7% explained)
Optimal clusters K*	2 (ARI stability = 1.000)
Livable countries	133 (76%)
XGBoost test F1	0.9455
XGBoost test AUC-ROC	0.9861
ABR-SHAP significant features	9 / 19
Overstatement Ratio (OR)	2.11
CMCS observed	0.400 (p = 0.0002, d = 14.43)
Noise stability (overlap)	>82% across all σ levels
Counterfactual feasibility	12.5% (1 of 8 non-livable test countries)

---

🛠️ Implementation Details

Component	Library / Version
Preprocessing & Clustering	scikit-learn 1.2
Classification	xgboost 1.7
Explainability	shap 0.41
Numerical Computing	numpy, scipy
Data Handling	pandas
Visualization	matplotlib, seaborn
Language	Python 3.9

Hardware: Intel Core i7-12700H · 32 GB RAM
Runtime: ~15–20 min (full ABR-SHAP B_max=500 + CMCS N_perm=5000)

---

🚀 Quickstart

# 1. Clone the repository
git clone https://github.com/<your-username>/<repo-name>.git
cd <repo-name>

# 2. Install dependencies
pip install numpy pandas matplotlib seaborn scipy scikit-learn xgboost shap tqdm

# 3. Update the data path in Cell 2
#    Change the read_csv path to your local CSV location

# 4. Run all cells sequentially in Jupyter
jupyter notebook Environment_Project.ipynb

Note: Cells must be executed in order (1 → 30) as each phase depends on outputs from the previous phase.

---

📋 Dataset

Source: country_environmental_livability_2019.csv
Scope: 175 country-level observations, year 2019
Feature domains:

Domain	Features
Environmental	CO₂ per capita, renewable energy %, forest area %, fossil fuel %, energy intensity, carbon damage
Economic	GDP per capita, manufacturing %, high-tech exports, trade openness, FDI inflow, R&D expenditure, air transport
Social	Life expectancy, electricity access, urban population %, fertilizer consumption, agricultural land %
Governance	Regulatory quality, control of corruption, government effectiveness

All percentage attributes validated within [0, 100]. No missing values across all 175 × 24 entries.

---

📄 Citation

If you use this code or methodology, please cite:

@article{rupok2026environmental,
  title     = {Environmental Livability Assessment via Adaptive Bootstrap-Retrained SHAP
               and Statistically-Constrained Pareto Counterfactuals: A Cross-National Analysis},
  author    = {Rupok, Junaid Ahmed},
  journal   = {IEEE Journal},
  year      = {2026}
}

---

📬 Contact

Junaid Ahmed Rupok
For questions regarding the methodology or implementation, please open an issue in this repository.