Dianatovar27/pump_it_up_ml_model

Machine learning model to predict the operational status of waterpoints in Tanzania. Includes data cleaning, feature engineering, EDA, and classification techniques to support infrastructure decision-making. # Pump it Up: ML Model to Predict Water Pump Status in Tanzania Machine learning model to predict the operational status of waterpoints in Tanzania. Includes data cleaning, feature engineering, EDA, and classification techniques to support infrastructure decision-making. Pump It Up - ML Model Predicting the functionality status of waterpoints in Tanzania using machine learning. 📌 **Objective** Classify each water pump into one of three categories: functional non functional functional needs repair 📂 **Dataset Description** The dataset contains information on over 59,000 waterpoints, including: Geographic location (latitude, longitude, region, district, etc.) Technical attributes (installer, funder, construction year, water quality, etc.) Operational information (permit status, extraction type, management type) Target variable: status_group ## 🚰 Workflow Overview 🧹 **Data Cleaning** Removed columns with only one value (recorded_by) or redundant information (payment_type, quantity_group, etc.). Handled missing values using: "Unknown" for categorical columns such as funder and installer. Mode imputation by group for scheme_management, grouped by management. Dropped columns that were too specific or had high cardinality (scheme_name, subvillage, wpt_name). 🧪**Feature Engineering** Applied transformations: Used np.log1p() on skewed numerical variables: population amount_tsh gps_height Created binary flags: is_amount_tsh_zero is_gps_height_zero has_valid_coords (when longitude > 1) has_valid_year (when construction_year >= 1960) Grouped infrequent categories into "Rare" using tailored frequency thresholds per variable. Reduced multicollinearity by removing raw variables like gps_height, population, and amount_tsh, while keeping only their log transformations and flags. Tested K-Means clustering on geolocation (latitude, longitude) — did not improve the model. Applied OptimalBinning on log_population — also did not improve results. 📊 * …