Accepted at IEEE International Conference on Data Mining (ICDM) 2025
DOI and publication link coming soon
Implementation of the horizon martingale framework for early change detection in dynamic networks.
This repository implements horizon martingales that leverage statistical forecasting to detect structural changes in dynamic networks before they fully manifest. The framework achieves 13-25% reduction in detection delay while maintaining rigorous false alarm control. For complete technical details, please refer to the documentation file.
Key Features:
- Early detection through predictive evidence accumulation
- Exact feature attribution via Martingale-Shapley equivalence
- Multiple network models (SBM, BA, ER, NWS) and real-world data support
git clone https://github.com/ali-izhar/anomaly_detection.git
cd anomaly_detection
pip install -r requirements.txt
# Run basic detection
python src/run.py -c src/configs/algorithm.yaml
# Run with custom parameters
python src/run.py -c src/configs/algorithm.yaml --network sbm --threshold 50The framework demonstrates consistent early detection across network types:
Horizon martingales (orange) achieve 22% delay reduction over traditional martingales (blue) for academic event detection, with TPR improvement from 88% to 96%.
| SBM (36% faster) | ER (81% faster) |
|---|---|
 |
 |
| BA (46%/17% faster) | NWS (46% faster) |
 |
 |
Horizon martingales consistently outperform traditional detection across all network types.
Demonstration of perfect equivalence (R² = 1.0000) between SHAP values and martingale values, with spectral features contributing 53.6% to detection.
@inproceedings{ali2025horizon,
title={Early Detection and Attribution of Structural Changes in Dynamic Networks},
author={Ali, Izhar and Ho, Shen-Shyang},
booktitle={IEEE International Conference on Data Mining (ICDM)},
year={2025}
}MIT License - see LICENSE for details.