From Zero to Hero: Cold-Start Anomaly Detection

Official PyTorch Implementation of "From Zero to Hero: Cold-Start Anomaly Detection" (ACL 2024).

Cold-Start Anomaly Detection

The cold-start setting provides two types of guidance:

1. A textual description of each normal class, serving as initial guidance, such as predefined topic names in chatbot systems.
2. 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

ColdFusion effectively adapts the zero-shot anomaly detector to contaminated observations.

Installation

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

Usage

2.1 Feature Extraction

python extract_features.py [--dataset] [--model]

2.2 Evaluation

Finally, you can evaluate by running the following command:

python coldfusion.py [--dataset] [--model]

Citation

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}
}