Predicting Severity of Traffic Accidents in the US

CIS 2450 Final Project — Erk & Brian

Setup

python3 -m venv venv
source venv/bin/activate     # macOS/Linux
# venv\Scripts\activate      # Windows

pip install -r requirements.txt

Drop accidentsData.csv into data/raw/ (gitignored — too large to commit).

Open notebooks/accident_severity_prediction.ipynb in VS Code or Jupyter.

Structure

├── data/
│   ├── raw/              # original CSV (gitignored)
│   └── processed/        # cleaned outputs (gitignored)
├── models/               # saved sklearn pipelines (.joblib)
├── modals/               # Dash dashboard components
├── notebooks/
│   └── accident_severity_prediction.ipynb   # full pipeline: EDA → 3 models
├── scripts/
│   └── balance_dataset.py          # one-time stratified sampling from raw 3 GB file
├── dashboard.py
├── requirements.txt
└── README.md

Dashboard Demo

Watch the demo on Loom

Presentation Video

[Watch the presentation on YouTube] https://www.youtube.com/watch?v=U7Awz0E0qz8

Difficulty Concepts

Four concepts applied; the three graded ones are marked ✅. Full justification and results are in the notebook sections linked below.

#	Concept	Notebook Section	Key evidence
✅ 1	Feature Importance	§6.1	RF MDI ranking — Hour and weather variables dominate; directly informs deployment recommendation in §8
✅ 2	Imbalanced Data	§3.2, §4 step 2	Downsampling to training set only after the split; Severe recall improved from ~0% baseline to 0.75
✅ 3	Hyperparameter Tuning	§5.1, §6.1, §7.1	RandomizedSearchCV over continuous distributions for GBM (finds between-grid-point optima); GridSearchCV for LR and RF
➕ 4	Ensemble Models	§6, §7	RF (bagging) + GBM (sequential residual correction) — two architecturally distinct ensemble strategies, each justified against the prior model's specific limitation

---

Application of Course Topics

This project applies the following course topics:

1. Polars: Used for efficient loading of the large accident dataset.
2. SQL / DuckDB: Used for SQL-based EDA queries on accident severity patterns.
3. Joins: Accident data was joined with Census population data to evaluate whether state-level severity patterns were confounded by population.
4. Hypothesis Testing: Chi-square, Mann-Whitney U, and Kruskal-Wallis tests were used to validate EDA findings.
5. Supervised Learning Models: Logistic Regression, Random Forest, and Gradient Boosting were trained for binary classification.
6. Different Hyperparameter Tuning Methods: GridSearchCV and RandomizedSearchCV were used to tune model parameters.
7. Time-Series / Temporal Features: Hour and day-of-week were extracted from accident start time.

---

Key Results

Model	Macro F1	Severe Recall	Notes
Logistic Regression (baseline)	0.678	0.66	Linear boundary; C=10 via GridSearchCV
Random Forest	0.722	0.75	max_depth=30 via GridSearchCV; recommended for deployment
Gradient Boosting	~0.720	~0.72	RandomizedSearchCV over 3 parameters

Recommended model: Random Forest — highest Severe recall (0.75), interpretable feature importances, and competitive macro F1.