README.en.md

Zero-Inflated Model for Accident Analysis

This project trains a Zero-Inflated Negative Binomial (ZINB) deep learning model on traffic accident count data before and after drowsy rest area installation, and analyzes influential variables using permutation importance.

Public Release Scope

• Raw data files (Excel), notebook outputs, draft documents (.docx), and archives (.zip) are excluded from Git tracking by default.
• Only executable code, research summary documents, and curated result CSV files are included for public release.

Folder Structure

.
├─ train_model.py
├─ requirements.txt
├─ .env.example
├─ results/
│  ├─ feature_importance_VERSION1.csv
│  └─ feature_importance_VERSION2.csv
├─ docs/
│  ├─ research_summary.md
│  └─ research_summary.en.md
├─ data/
│  ├─ README.md
│  └─ README.en.md
└─ .gitignore

How to Run

1. Install dependencies

pip install -r requirements.txt

2. Run training

python train_model.py --data-path "<your_excel_path>" --output-dir results

You can also pass the path through the DATA_FILE environment variable (see .env.example).

Optional arguments:

• --versions VERSION1 VERSION2
• --epochs 500
• --learning-rate 0.001

Key Results (Summary)

• Top variables in Version1: 연속주행시간, 통행속도, 2시간 이상 주행차량 대수
• Top variables in Version2: 톨게이트(개수), 통행속도, 차로수
• Is_After importance is low or negative in both versions, suggesting road/traffic features contributed more strongly than the before/after indicator itself.

For details, see docs/research_summary.md and docs/research_summary.en.md.

Example: Upload to a Separate GitHub Repository

git init
git add .
git commit -m "Initial public release"

After connecting a remote repository:

git branch -M main
git remote add origin <your_repo_url>
git push -u origin main