the definitions of unsupervised and semi-supervised approaches

[Pengyu-Li-2002]

Dear Author,

Thank you for your open-source work; it has significantly advanced the field of time series anomaly detection (TSAD). However, I have some questions. In the code, you use the training set for both training and validation, and the test set solely for inference, considering this approach as a semi-supervised method. Does this differ from the current definitions of unsupervised and semi-supervised approaches commonly used in the TSAD field? I look forward to your response.

Thank you!

[qhliu26]

Hello Pengyu,

Thank you for your interest in our work! I think this is a common practice in this field. Actually, the test set is the entire time series, and the training set is the initial portion of the time series which exhibits no anomaly or very few anomalies.

Here I attach some definitions of unsupervised/semi-supervised approaches in TSAD for reference: (i) unsupervised, which does not require any labeled data. So we directly pass the test set into the anomaly detectors. (ii) semi-supervised, requiring labels only for normal instances. So, we use the training set to fix/train the model before applying it to the test set.

Let me know if I address your questions and feel free to offer suggestions!