Live demo: https://undp-sdg-analytics-5jmhuht64bmnybgwmzpkj9.streamlit.app

🌍 UNDP SDG Analytics Dashboard

An interactive data science dashboard analyzing UN Sustainable Development Goal (SDG) indicators across 24 countries and 6 regions — built while preparing for the UNDP Digital, AI and Innovation Hub Internship.

🔍 What It Does

• K-Means Clustering — groups countries into Vulnerable / Developing / Advanced tiers based on HDI, poverty rate, internet access, and CO₂ per capita
• Elbow Method + Silhouette Score — statistically evaluates candidate values of k rather than picking one arbitrarily
• Sensitivity Analysis (k=2–6) — compares k=3 against every other reasonable choice on silhouette score, including k=4 as a documented alternative, instead of asserting k=3 is automatically optimal
• Missing Data Handling — the raw dataset includes realistic missing values; cleaning uses regional median imputation, with the before/after shown transparently in the app
• Gini Coefficient — quantifies income inequality across GNI per capita
• HDI Trend Analysis (2015–2023) — time-series view capturing the COVID-19 regression visible in several countries' HDI scores

📊 Data Sources

Indicator	Source
Human Development Index (HDI)	UNDP Human Development Report 2023
Poverty rate, literacy, internet access, GNI/capita	World Bank World Development Indicators
CO₂ per capita	Our World in Data

Countries were manually curated across 6 regions (Sub-Saharan Africa, South Asia, Latin America, Europe, East Asia & Pacific, MENA) to ensure geographic diversity rather than selecting only data-complete or high-income nations.

🛠️ Tech Stack

• Python — core language
• Pandas / NumPy — data loading, cleaning, transformation
• scikit-learn — K-Means clustering, MinMax scaling, silhouette scoring
• Plotly — interactive charts (scatter, line, bar)
• Streamlit — dashboard framework

🧹 Data Cleaning Approach

Real SDG data is rarely complete. data/sdg_indicators_raw.csv intentionally includes missing values (as World Bank/UNDP exports often do) to demonstrate the cleaning pipeline in app.py:

1. Missing numeric values are imputed using the median of other countries in the same region — not a global median — since development indicators correlate strongly within a region.
2. If an entire region's values are missing for a column, a global median fallback is used.
3. The "Data Cleaning" tab in the app shows the missing-value report and a before/after comparison so the imputation is auditable, not hidden.

🔵 Why k=3 for Clustering — and Why Not k=4

k=3 is used as the primary model throughout this dashboard, but it is not the only value tested, and it is not automatically the "best" one by the numbers alone. The app includes a full Sensitivity Analysis (k=2–6) tab that runs K-Means across that whole range and reports silhouette scores honestly:

k	Silhouette Score
2	0.478
3	0.460
4	0.487
5	0.505 (highest)
6	0.481

k=3 does not have the highest silhouette score in this dataset — k=5 does. That's stated plainly in the app rather than hidden, because pretending the chosen k was automatically optimal would misrepresent what the statistics actually show.

So why keep k=3 as the primary model?

• It maps directly onto UNDP's own Low / Medium / High human development framework, which makes the clusters interpretable for policy discussion — a property no purely statistical metric captures.
• The silhouette differences across k=2 to k=6 are small (0.46–0.51), so no single k is decisively "correct" on separation quality alone. In that regime, interpretability is a reasonable tiebreaker.
• At k=4 specifically, the dashboard shows a country-by-country comparison: the extra cluster splits the existing Vulnerable group into two sub-groups, while Developing and Advanced stay intact. k=4 adds resolution at the low-HDI end rather than telling a different story — it's documented in the app as a legitimate alternative, not dismissed.

All four clustering features (HDI, poverty rate, internet access, CO₂ per capita) are Min-Max normalized to [0,1] before clustering — without this, GNI-scale features would dominate the Euclidean distance calculation purely due to unit scale, not real similarity. (GNI itself is excluded from the clustering features to avoid the clusters simply re-deriving income brackets — it's examined separately via the Gini coefficient instead.)

All four clustering features (HDI, poverty rate, internet access, CO₂ per capita) are Min-Max normalized to [0,1] before clustering — without this, GNI-scale features would dominate the Euclidean distance calculation purely due to unit scale, not real similarity.

🎯 SDGs Covered

• SDG 1 — No Poverty
• SDG 4 — Quality Education (literacy)
• SDG 9 — Industry, Innovation & Infrastructure (internet access)
• SDG 10 — Reduced Inequalities (Gini coefficient)

🚀 Running Locally

git clone https://github.com/<your-username>/UNDP-SDG-Analytics-Dashboard.git
cd UNDP-SDG-Analytics-Dashboard
pip install -r requirements.txt
streamlit run app.py

The app will open at http://localhost:8501.

📁 Project Structure

UNDP-SDG-Analytics-Dashboard/
│
├── app.py                        # Streamlit app: cleaning, clustering, visualization
├── utils/
│   ├── data_loader.py            # Data ingestion, validation, cleaning pipeline
│   ├── clustering.py             # K-Means, Gini coefficient, elbow method — from scratch
│   └── charts.py                 # Plotly chart-building functions
├── notebooks/
│   └── EDA.ipynb                 # Exploratory data analysis walkthrough
├── data/
│   ├── sdg_indicators_raw.csv    # Raw data with realistic missing values
│   ├── sdg_indicators.csv        # Reference clean dataset
│   ├── hdi_trends.csv            # 2015–2023 HDI time series
│   └── data_sources.md           # Column-level source documentation
├── requirements.txt
├── README.md
└── screenshots/
    ├── dashboard.png
    └── analysis.png

app.py contains a self-contained version of the pipeline for the live Streamlit dashboard. utils/ holds the same logic as reusable, testable modules (including a from-scratch K-Means implementation), and notebooks/EDA.ipynb walks through the analysis step by step — useful for showing your reasoning in an interview.

Development Approach

This project was built with AI-assisted development (Claude, Anthropic) for code scaffolding. My own contributions: problem framing and SDG indicator selection, data source research, methodology decisions (clustering features, k selection, imputation strategy), and the analysis and policy interpretation throughout. AI-assisted development is standard practice in modern data science workflows, and this project reflects that approach while keeping the methodology fully explainable and defensible.

📄 License

MIT — feel free to fork and extend with more countries or indicators.