saky-semicolon/Zindi_African-Credit-Scoring-Challenge

Zindi Competition: African Credit Scoring Challenge, Predicting loan default probabilities using machine learning with feature engineering, model optimization, and performance evaluation. # African Credit Scoring Challenge ## Project Overview This repository contains the implementation for the **African Credit Scoring Challenge** hosted on Zindi. The challenge involved predicting the likelihood of a customer defaulting on a loan based on financial data. The project utilizes **deep learning techniques** combined with **feature engineering** and **hyperparameter optimization** to achieve high predictive performance. ## Dataset The dataset consists of: - **Training Data**: Customer financial records with features like loan type, demographics, and repayment history. - **Test Data**: Customer records without target labels, used for final evaluation. - **Target Variable**: A binary classification label indicating **loan default (1) or no default (0).** ## Key Features & Methodology ### 1. Data Preprocessing - **Handling Missing Values:** - Numerical features were imputed with the mean. - Categorical features were imputed with the mode. - **Feature Engineering:** - `days_to_repay`: Difference between loan disbursement date and due date. - `repay_ratio`: Ratio of total repayment amount to original loan amount. - `amount_duration_ratio`: Ratio of loan amount to loan duration. - `disbursement_month`: Extracted from the disbursement date. - Interaction terms such as `amount_duration_interaction` and `repay_duration_interaction`. - **Categorical Encoding:** - Label encoding for categorical variables such as `loan_type` and `country_id`. - **Normalization:** - Standardization using `StandardScaler` to ensure zero mean and unit variance. - **Class Balancing:** - **ADASYN (Adaptive Synthetic Sampling)** was used to balance class distribution. - **Data Splitting:** - 80-20 split into training and validation sets. ### 2. Deep Learning Model The architecture is based on a **Wide and Deep Learning** framework: - **Wide Layer**: Captures simple feature interactions. - **Deep Layers**: - **LayerNorm** for stable training. - **LeakyReLU** activation. - **Dropout** for …