Zurich Traffic Accident Severity Prediction: Analysis of a Data Science Fundamentals Workshop Project

This project is a practical outcome of the Data Science Fundamentals Workshop. It uses machine learning techniques to predict the severity of traffic accidents in Zurich, Switzerland, demonstrating the complete data science workflow from data exploration to model deployment. It covers key stages such as problem definition, data processing, model construction, and evaluation, and has practical application value for traffic safety management.

Traffic accidents are a major global public health challenge. According to the World Health Organization, approximately 1.3 million people die from road traffic accidents each year. Accurately predicting accident severity is of great significance for optimizing emergency resource allocation, improving road design, and formulating insurance strategies. As Switzerland's largest city, Zurich has well-developed traffic infrastructure and open traffic accident data, providing valuable practical materials for this project.

The project follows the standard data science lifecycle:

1. Problem Definition: Predicting accident severity, which is a supervised learning classification/regression problem;
2. Data Acquisition and Exploration: Using Zurich's open data, which includes dimensions such as basic accident information, road characteristics, and participant information. It is necessary to analyze data distribution, correlation, and quality;
3. Preprocessing and Feature Engineering: Cleaning (handling missing/anomalous values), encoding (categorical variables, time/geographic features), and constructing combined/aggregated/ratio features;
4. Model Selection and Training: Trying traditional models such as logistic regression, random forest, and gradient boosting. Deep learning models can be used when data is sufficient;
5. Model Evaluation: Using metrics such as accuracy, precision/recall, F1 score, ROC-AUC, and confusion matrix. For safety-critical applications, more attention is paid to the recall rate of severe accidents;
6. Result Interpretation: Mining insights such as influencing factors and high-risk scenarios.

1. Class Imbalance: Severe accidents account for a low proportion. Solutions include resampling (SMOTE oversampling, undersampling), cost-sensitive learning, and threshold adjustment;
2. Feature Importance: Using tree model feature importance and SHAP values to explain model decisions;
3. Spatiotemporal Patterns: Mining time (peak hours, holidays), space (intersections, road sections), and interaction effects (bad weather + night + high speed);
4. Model Generalization: Avoiding overfitting through cross-validation, time series segmentation, and independent dataset validation.

Potential Applications:

1. Pre-deployment of emergency resources;
2. Insurance pricing;
3. Road safety audits;
4. Driving behavior intervention. Limitations:
5. Data bias (reflects past road conditions and law enforcement standards);
6. Correlation ≠ Causality;
7. Ethical considerations (avoiding discriminatory treatment);
8. Privacy protection (risks of location data).

For Beginners: Provides learning materials for experiencing the complete workflow, facing real data challenges, integrating domain knowledge, and understanding result interpretability. For Urban Planners and Traffic Managers: The model can serve as a decision support tool, but it needs to be used in combination with traditional professional knowledge.

This project is a typical entry-level data science project, demonstrating the application potential of machine learning in public safety management. By predicting traffic accident severity, it not only practices the data science technical workflow but also provides a data-driven perspective for improving urban traffic safety, proving that data science is a tool for solving practical social problems and creating public value.