[ENH] Implement Extended Isolation Forest

[SebastianSchmidl]

Describe the feature or idea you want to propose

The Extended Isolation Forest improves upon the Isolation Forest method.

Describe your proposed solution

Implement directly in aeon:

• Paper: https://doi.org/10.1109/TKDE.2019.2947676
• Code: https://github.com/sahandha/eif
• Code (in TimeEval): https://github.com/TimeEval/TimeEval-algorithms/tree/main/eif

However, support semi-supervised as well as unsupervised usage similar to the PyODAdapter or KMeansAD

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[Akhil-Jasson]

Hi, I came across this issue and found it really interesting. I'm not entirely sure if I can solve it, but I'd love to give it a try. Would it be okay for me to work on this? Also, if there are any specific guidelines or things I should keep in mind while implementing this in Aeon, I'd appreciate the guidance.

[Akhil-Jasson]

When implementing the EIF, am I supposed to use aeon's Base Anomaly Detection?

[baraline]

Yes, you should use the base class of the module you are implementing the estimator to

[Akhil-Jasson]

I'm encountering issues while importing the EIF model referenced above. When attempting to install the package, I receive a ModuleNotFoundError for numpy, even though I have already installed numpy in my environment. Could anyone suggest a fix? Alternatively, can I use h2o.ai's EIF model as a viable alternative if this installation problem persists?

[MatthewMiddlehurst]

What are you trying to install? We would need more information to help. Look through our contributor guide on the website if you have not already.

[Akhil-Jasson]

@MatthewMiddlehurst I'm trying to install the eif package mentioned in the issue above (https://github.com/sahandha/eif), but I'm facing the above difficulty.

[MatthewMiddlehurst]

This package has not been updated in 5 years. We generally prefer direct implementations rather than wrappers where possible. In this case I would reject code wrapping the package due to its unmaintained state. I believe the package was just linked as an example of an implementation.

[Akhil-Jasson]

@MatthewMiddlehurst Since the eif package appears to be unmaintained, would using H2O.ai's implementation of EIF (https://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/eif.html) be a viable alternative? Let me know if there are any preferred approaches for this use case

[MatthewMiddlehurst]

I have never used or heard of H2O.ai sorry. A wrapper of another package is better than nothing I guess, but the best option would be to have our own implementation. This algorithm seems to be widely available.

[Akhil-Jasson]

Please let me know if there's anything that needs to be modified or improved.