[QUESTION] Have you considered more advanced outlier detection and value imputation?

[reiser02]

Hi,

I am working with low-quality time series from low-quality sensors in cities, and the problem of detecting outliers and imputing values to obtain higher-quality time series is a major concern.

I’d like to know if you’ve considered this.

Here’s a model that I find very interesting: https://huggingface.co/ibm-granite/granite-timeseries-tspulse-r1

[dennisbader]

Hi @reiser02, thanks for this feature request and sorry for the late response, we were wrapping up some features for Darts :)

Foundation models for anomaly detection and imputation could indeed be interesting to add. It's not on one of our top priorities at the moment though, but I'm very open to add it.

I could imagine something like this:

• A new FoundationAnomalyModel (or similar nname): This would go into our anomaly_detection module. Depending on what the output of the foundation model is, it could either be a subclass of AnomalyScorer or AnomalyDetector. We should leverage our existing FoundationModel class which already implements the torch and foundation model (e.g. hugging face connector) logic.
• A new FoundationImputer (or similar name): This would go into our filtering module. It should be a subclass of FilteringModel . The filter() method would then fill the missing values on series' history. Same as for the FoundationAnomalyModel, we should leverage the existing FoundationModel.

Any thoughts on this?