Natural Disaster Severity Prediction

This repository contains the code and reproducible submission pipeline for the Data Mining Spring 2026 final project.

Final Result

Best verified public leaderboard score:

Public MAE: 0.8151

Final submission file:

submissions/final_submission.csv

The final submission is copied from:

submissions/c28_after8162_C28_HYBRID_G275_R120_CAP840.csv

Method Summary

The final system uses deterministic ensemble calibration and post-processing. The main stages are:

1. Strict 91-day based learned-rank signal generation.
2. Distribution-preserving calibration of prediction scores.
3. Public-anchor guided continuation using verified submissions.
4. Hybrid calibration combining validated continuation and learned rank agreement.

No test labels, private leaderboard labels, or external datasets are used.

Environment Setup

python -m venv .venv ..venv\Scripts\Activate.ps1 pip install -r requirements.txt

Data Preparation

Place the competition files in:

data/train.csv data/test.csv data/sample_submission.csv

The raw data files are not included in this repository.

Public Score Log

The public leaderboard progress is recorded in:

outputs/public_score_log.csv

Reproducibility

All final scripts are deterministic. The post-processing scripts preserve the official sample submission order and required columns. Predictions are clipped to the valid score range.