Official PyTorch Implementation of "From Zero to Hero: Cold-Start Anomaly Detection" (ACL 2024).
The cold-start setting provides two types of guidance:
- A textual description of each normal class, serving as initial guidance, such as predefined topic names in chatbot systems.
- A stream of t contaminated observations (that may include anomalies), e.g., real user queries.
It is particularly relevant in real-world applications where, shortly after deployment, a short stream of user queries becomes available but the queries are not labeled into intent types and some of them are out-of-scope.
ColdFusion effectively adapts the zero-shot anomaly detector to contaminated observations.
Create a virtual environment, activate it and install the requirements file:
virtualenv -p /usr/bin/python3 venv
source venv/bin/activate
pip install -r requirements.txt
python extract_features.py [--dataset] [--model]
Finally, you can evaluate by running the following command:
python coldfusion.py [--dataset] [--model]
If you find this useful, please cite our paper:
@article{reiss2024zero,
title={From Zero to Hero: Cold-Start Anomaly Detection},
author={Reiss, Tal and Kour, George and Zwerdling, Naama and Anaby-Tavor, Ateret and Hoshen, Yedid},
journal={arXiv preprint arXiv:2405.20341},
year={2024}
}