Tanzanian Waterwells Project

Overview

The Tanzanian Waterwells Project is an initiative aimed at predicting water pump functionality to ensure clean water access for the people of Tanzania. This project utilizes machine learning (ML) techniques to classify water wells as functional, non-functional, or in need of repair. By leveraging data analysis and model optimization, our goal is to provide the people of Tanzania a ML model that can make accurate predictions on water pumps that are functional and produce clean water.

Team

• Shefat Moral
• Gabriel Santorelli
• David Jimenez

Project Goals

• Our goal for this project was to build an effective ML model that can predict water pump functionality for the purpose of bringing clean water to the people of Tanzania 🇹🇿.

---

Data Exploration (EDA)

Key Findings:

• The dataset contained three well statuses:
  • Functional
  • Non-Functional
  • Functional Needs Repair
• Functionality by Region: Analyzed geographic patterns of well status.
• Water Quality Analysis: Functioning wells did not always provide clean water.
  • Categories like Unknown, Fluoride Abandoned, and Salty were among the lowest quality.

---

Model Development

Baseline Model: Decision Tree

• Initial accuracy: 75%
• Strengths: Identified Class 0 (Functional) and Class 2 (Non-Functional) well.
• Weakness: Struggled with Class 1 (Needs Repair).
• Class Balancing: Addressed class imbalance with class_weight and SMOTE (Synthetic Minority Over-sampling Technique).

Improved Model: Random Forest

• Initial accuracy: 79%
• Improved precision for Class 1 (60%) and Class 2 (84%).
• Struggled to distinguish Class 1 from Class 2.
• Further EDA to determine the next steps in optimization.

Feature Engineering & Model Optimization

• Correlating Features: Engineered features based on relationships (e.g., gps_height, function_years).
• Hyperparameter Tuning: Adjusted parameters for better classification performance.
• Error Analysis: looked into the features that were potentially causing false positives.

Final Model Performance

• Optimized Random Forest Model
  • Error Analysis: Pumps that needed repair weren't producing good water. Grouped Needs Repair with Not functional.
  • Accuracy: up to 81%, that's a 6% increase from our baseline model(75%).
  • Precision: up to .82, that's a 36% increase from our baseline model(.46).
  • Recall: up to .74, that's a 42% increase from our baseline model(.32).
  • F1-Score: up to .78, that's a 40% increase from our baseline model(.38).
  • Reduced false positives by 200 cases.
  • Mean CV Accuracy: .793, indicating there is no overfitting or data leakage.
  • Standard deviation: of 0.0025, indicating model stability.

---

Future Improvements

• Continue Error Analysis: Keep exploring features that are causing false positives.
• Explore additional models: Like SVM, Gradient Boosting, etc., to enhance the accuracy of the model further.

Acknowledgments

• We appreciate the Tanzanian government and local organizations for bringing clean water to communities. Special thanks to our fellow peers and instructor for their guidance throughout this project.