[ENH] Base class for whole-series (collection) anomaly detection

[MatthewMiddlehurst]

Describe the feature or idea you want to propose

While the primary goal of our anomaly detection (AD) module is to detect anomalous points or subseries in single time series, there is also a task for whole-series AD. In whole-series AD you have a collection of series, and the goal is to select the label the series which are anomalous. This can be similar to rare-class classification in some cases, but there are also unsupervised methods.

Describe your proposed solution

Implement framework for whole-series anomaly detection as a submodule of anomaly_detection. This includes a base class extending from BaseCollectionEstimator.

Describe alternatives you've considered, if relevant

I do not believe this is large (or different) enough to warrant its own module. An AD submodule seems like the correct way to start this.

Additional context

No response

[MatthewMiddlehurst]

See #2326 for draft

[KowshikB03]

Is there a specific reason why Isolation Forest is used in #2326 draft? Should we experiment with other classifiers?

[MatthewMiddlehurst]

Its just an example estimator