Quick start, complete sample code which can generate the results in the picture.

[943fansi]

Quick start, complete sample code which can generate the results in the picture.

[angeliney]

Hi @943fansi, thank you for your interest and sorry for the wait! We concluded that from our experience playing with toy data and open source datasets. But, I would suggest to definitely verify that information yourself, since it maybe different for different datasets. For training set, we suggest to just use normal data in the case of prediction and reconstruction method, especially because the models are first trying to learn what normal looks like and we're using the predictive or reconstruction error as a gauge of abnormality.

[benjaminye]

Hi @943fansi! Thanks for checking out this project.

You can learn more about time series anomaly detection through our blog posts.
Blog #1 deals with use cases.
Blog #2 outlines different techniques.

And most importantly:
Blog #3 gives you concrete example on how to use the package. The notebook is included in the blog post.

I will close this issue once I incorporate the contents into the documentation.
Sorry for the inconvenience.

Thanks again!
Ben

[943fansi]

Hi, @benjaminye. Thank you very much for your reply. The article is very good. I have some questions about anomaly detection that I would like to ask you. "In general, predictive methods work well for point similarities while reconstruction and distance-based methods work well for pattern similarities." Where does this conclusion come from? I am studying anomaly detection based on prediction and reconstruction methods. I am confused about the training set. Should I only use normal data or include abnormal data?
Look forward to your favorable reply!

[943fansi]

@angeliney Thanks.