LittleDarkBug/2024_togo_fiber_optics_prediction_challenge_submission_zindi

Final submission for the togo fiber optics adoption prediction challenge on Zindi. 6th place solution # Fiber Optics Adoption Prediction Notebook ## Overview and objectives The goal of this notebook is to predict fiber optics adoption based on categorical data and geo data from MOSAIKS API. ### Expectations A model that generalize optimally well on unknown data, and have less train and inference time. ### File Description This file is responsible for preprocessing the dataset, selecting meaningful geo features, constructing a model, evaluating the model, and creating a submission. ## Performance metric The Evaluation Metric is AUC ### Score on Zindi Private : `0.92589805` Public : `0.922673591` ## ETL process ### Data Used Train DataSet given in the competition `(Train.csv)` Test DataSet given in the competition for submission creation `(Test.csv)` ### Features Used The features are split in two. The categorical one, and the geo data (Numerical). All Categorical feature are used in prediction, but only a handfull of numerical features are used. The choice of numerical feature is done by Elimination based on correlation matrix. The categorical Features Explanation are given in `Metadonnées.csv` file ### Input Files - The train set : stored as `Train.csv` in the same directory - The test set : stored as `Test.csv` in the same directory ## Data Modeling ### Description of the Data Model(s) The project utilizes a voting ensemble method combining predictions of several powerful classifiers: CatBoostClassifier, XGBoostClassifier, LightGBMClassifier, and AdaBoostClassifier. The voting approach leverages the strengths of each individual model to improve overall performance. This methodology is grounded in the theory that diverse models, when combined, can produce more robust and accurate predictions. #### Model architecture ### Feature Selection, Engineering, and Normalization Feature selection was performed by eliminating multicollinearity among numerical features. Multicollinearity can distort the performance of machine learning models, leading to …