Traffic Crash Cause Identification

Multi-class ML pipeline analyzing Chicago crash records.

A large-scale machine learning project focused on identifying contributory causes of traffic accidents. The pipeline involved massive data consolidation, feature engineering, and high-performance classification.

Technical Stack

  • Languages & Libraries: Python, Scikit-learn, Pandas, NumPy.
  • Models Tested: Decision Tree, Logistic Regression, Gaussian Naïve Bayes, Multinomial Naïve Bayes, kNN, Random Forest, SVM.
  • Evaluation: Custom multi-label framework (Precision, Recall, F1-score).

Core Features

  • Data Engineering: Consolidated 794k+ Chicago crash records, mapping 40 contributory causes into 5 actionable classes for simplified classification.

  • Feature Optimization: Streamlined the feature space from 48 to 18 key predictors, reducing noise and improving model computational efficiency.

  • Model Performance: Optimized a k-Nearest Neighbors (kNN) model to achieve 54.19% accuracy, a 15% improvement over the baseline and the highest among all tested classifiers.

Project Workflow

Below is the architectural flow of the ML pipeline: flowchart.png

Architecture of the data preprocessing and model selection pipeline.