[ENH] Implementation of Extended Isolation Forest (EIF) anomaly detector #2679
Conversation
Akhil-Jasson
requested review from
SebastianSchmidl and
MatthewMiddlehurst
as code owners
Thank you for contributing to
|
|
Please fill out the template and use the correct title format. |
MatthewMiddlehurst
added
the
enhancement
Akhil-Jasson
changed the title
|
I've fixed the title format. Shall I proceed with adding EIF to the online API documentation? |
|
Hi @Akhil-Jasson, I just saw your code, and it looks great! That said, I don’t think using H2O is the best approach since Aeon doesn’t rely on it. It might be better to have our own implementation instead. A few updates to consider: 1.Could you update the section "Does your contribution introduce a new dependency?" and mention H2O there? 2.The test cases seem to be missing—could you add them? 3.Instead of importing the entire H2O module, it’s better to import only what’s needed to keep things lightweight. |
|
New dependencies should be put in pyproject.toml otherwise this won't be tested, Still bits missing from the template |
…ile for the EIF implementation.
|
I've added the h2o dependency to pyproject.toml, but I'm encountering errors when running the test files. The test attempts to import aeon.anomaly_detection._eif but fails, indicating that the module doesn’t exist yet. Is there a step I'm missing for adding new modules to aeon? What could be the possible issue? |
|
your import is incorrect. |
There was a problem hiding this comment.
AI2O looks like a massive package. Do we actually want to include it as a dependency? The issue clearly states that we are looking for an implementation in aeon directly.
I know this is not mentioned in the corresponding issue, but I think it makes sense to work with sliding windows in EIF as well. We can always get the original behavior back by setting the window-size to 1.
| self.contamination = contamination | ||
| self.extension_level = extension_level | ||
| self.random_state = random_state | ||
| self.scaler = StandardScaler() |
There was a problem hiding this comment.
scaler is not a parameter, so cannot be instantiated here. Please read the sklearn estimator development docs on how to name and where to place parameters, fitted and non-fitted attributes, etc.
| # Fit the scaler | ||
| self.scaler.fit(X) | ||
| X_scaled = self.scaler.transform(X) |
There was a problem hiding this comment.
The paper does not explicitly mention that scaling the data provides better results. In anomaly detection scaling might hide some types of anomalies. Why do you include it?
|
|
||
| return self | ||
|
|
||
| def _predict(self, X) -> np.ndarray: |
There was a problem hiding this comment.
To make the usage of EIF similar to our other models, we want it to be usable as a semi-supervised (as implemented already) and an unsupervised algorithm. The current implementation of _predict does not allow that.
This PR implements the Extended Isolation Forest (EIF) algorithm.
Reference Issues/PRs
Fixes #2113
What does this implement/fix? Explain your changes.
Does your contribution introduce a new dependency? If yes, which one?
Yes, it introduces H20.ai as a new dependency
Any other comments?
PR checklist
For all contributions
For new estimators and functions
__maintainer__at the top of relevant files and want to be contacted regarding its maintenance. Unmaintained files may be removed. This is for the full file, and you should not add yourself if you are just making minor changes or do not want to help maintain its contents.For developers with write access