Sebastian Schmidl
Sebastian Schmidl
When reading the documentation of PyThresh, I noticed some issues. This PR fixes the following: - The PyDoc documentation of the `pythresh.thresholds.base.BaseThresholder` was not rendering correctly. - The threshold attribute...
I integrated PyThresh into our [TimeEval tool](https://github.com/HPI-Information-Systems/TimeEval/blob/main/timeeval/metrics/thresholding.py#L345) and I noticed that REGR builds a normal distribution without seeding the RNG. This leads to different results when calling it multiple times...
There are even more datasets for time series anomaly detection. We should add them to our collection: 1. Inspect dataset description if it fits our task definition. 2. Download and...
We introduced 3 global phases for the execution of the evaluation. The `DockerAdapter` already uses the first phase "prepare" to download the images. However, the other adapters were not updated....
PR #115 introduces a scheduled workflow that automatically creates a PR with the updated algorithm stubs. Example: #116. However, the (typecheck and test) workflows for this PR are not triggered....
When you install TimeEval via `pip install timeeval`, the examples scripts in the documentation do not work out-of-the-box because the datasets are missing. We should package some example datasets within...
Currently, just using external datasets is cumbersome in TimeEval. Ideas for improvement: - Add an implementation of the `Datasets`-interface just for custom datasets: `CustomDatasetManager` - Allow specifying the dataset information...
The documentation on using external datasets with TimeEval is not complete and not consistent: - Fix code samples (remove HPI constants) - Add a section on how to create your...
Integrate the data, metrics, and methods of the TimeSeAD library into TimeEval: - Paper: https://ml.informatik.uni-kl.de/publications/2023/TimeSeAD.pdf - Code: https://github.com/wagner-d/TimeSeAD Changes: - [ ] Adjusted SMD collection datasets with only 15 time...
Many deep learning papers use the _point-adjusted F1_ measure ($F1_{PA}$) as a quality criterion, despite that it is highly susceptible to overestimating the detection performance [1]. We should include this...