Work / Case Study
Traffic Collision Prediction
Machine Learning for Urban Collision Severity
An end-to-end Python pipeline that turns NYC collision records into engineered features, compares ensemble models and generates evaluation reports for severity prediction.
From raw data to model insight
- NYC collision records
- Data & feature pipeline
- Ensemble models
- Evaluation & reports
The Research Problem
Raw urban collision records contain missing values, mixed data types and complex relationships. A useful modeling workflow needs consistent data preparation and a way to compare approaches before interpreting predictions.
The Solution
The repository separates data processing, model training and visualization into a repeatable pipeline. It prepares collision records, engineers features, compares ensemble classifiers and saves models alongside analysis reports.
Pipeline Architecture
The main stages of the research pipeline.
- 01NYC collision records
- 02Data preparation
- 03Feature engineering
- 04Ensemble training
- 05Evaluation
- 06Reports & saved model
Key Capabilities
Real-world Data
Download and prepare NYC Open Data collision records for analysis.
Feature Engineering
Transform raw collision data into features for model training.
Ensemble Models
Compare Random Forest, Gradient Boosting and Extra Trees.
Evaluation & Reports
Generate model comparisons, feature importance and visual analysis.
Engineering Focus
Reproducible pipeline
Separate data processing, model training and visualization so experiments can be run and inspected step by step.
Model comparison
Evaluate multiple ensemble classifiers using the same prepared data and generate comparison reports.
Interpretable outputs
Use feature-importance plots and classification reports to inspect model behavior beyond a single score.
The Result
A public research implementation covering the path from raw collision records to saved models and visual reports. The repository provides the code and setup instructions for reviewing the approach.
Technologies
Python · Pandas · NumPy · Scikit-learn · Matplotlib · Seaborn · Joblib