natalienyabanhi/road-accident-ml-project

Machine learning and reinforcement learning system for analysing and predicting South African road accident severity using ensemble models and Q-learning. #South Africa Road Accident Analysis & Intelligent Decision System (2017) Project Overview This project is a complete data science and machine learning system that analyzes road accident data in South Africa (2017). It includes data preprocessing, exploratory data analysis (EDA), ensemble machine learning models and a reinforcement learning (RL) system for accident prevention recommendations. The goal is to predict accident severity and recommend traffic interventions using AI techniques. ## Dataset Source The dataset used in this project is publicly available and contains road accident records from South Africa (2017). Source: kaggle.com ## Dataset Description The dataset contains 120 records and 16 features, including: - Accident severity - Location type - Province and city - Vehicle type - Speed and speed zone - Number of vehicles and casualties - Date and time of accident - Road conditions (Occasions) ## Technologies Used - Python - Pandas - NumPy - Matplotlib - Seaborn - Scikit-learn - XGBoost - CatBoost ## Data Preprocessing The dataset was cleaned and prepared using the following steps: - Artificial missing values (8%) introduced for simulation - Missing values handled using: - Mode imputation (categorical variables) - Median imputation (numerical variables) - Date and Time converted into usable formats - Outliers detected using IQR method and capped (Winsorisation) - Feature engineering: - Hour extracted from Time - Month extracted from Date - Label encoding applied to categorical variables - Target variable encoded: - 0 = Bumper Accident - 1 = Headon Accident - 2 = Fatal Accident --- ## Exploratory Data Analysis (EDA) EDA was performed to understand: - Accident severity distribution - High-risk provinces - Vehicle types involved - Casualty distribution patterns - Time and seasonal trends ## Machine Learning Models Three ensemble models were used: - Random Forest Classifier - XGBoost Classifier - CatBoost Classifier ### E …