whyyoutoo/Tanzanian-Water-Well-Functionality

Predicting functionality of water points in Tanzania and identify factors that lead to failure. # Tanzanian-Water-Pump-Functionality **Authors**: Wesley Yu ## Overview This project will use machine learning techniques to predict functionality of waterpoints in Tanzania. ## Business Problem The Tanzanian government has set aside a budget to repair non functional water wells. They would like a method to predict the functionality of water wells so that resources can be allocated for repairs. They would also like to know what factors lead to a waterpoint being non functional, so that newly installed waterpoints will have better chances of performing as needed. ## Data Data set taken from a Driven Data competition containing around 60,000 records of water points in Tanzania. Each record contains various information of the water points such as, type of pump, location, management, water source, and water quality as well as the functionality of the water point. Link to competition page: Pump it Up: Data Mining the Water Table ## Methods Various machine learning algorithms will be tested and hyperparameters tuned to find optimal results based on f1 score. ## Results Final model was a random forest classifier with tuned hyperparameters. It performed decently in predicting non functional waterpoints, with a precision of 85% and recall of 73%. Decision threshold can also be adjusted for higher precision at the cost of recall, depending on buisness needs. Permutation importance shows top 5 features with greatest effect are quantity, waterpoint_type, payment, extraction_type_group, and construction_bins. Older waterpoints were found to have more non functional than functional. Dry water quantity has the highest majority of non functional waterpoints. Waterpoints that do not require payment to use have the majority of non functional. This may be due to less resources avalible to maintain waterpoint. 4 out of the top 6 extraction type were different models of handpumps (nira/tanira, afridev, india mark ii, and swn 80). ## Conclusions Based …