Proposed addition: MemStream anomaly detection for River

[NicolasNoya]

Hi maintainers,

I would like to propose the addition of a MemStream implementation to the River library.

This work was carried out as part of a Data Streaming course, with the goal of implementing and integrating a state-of-the-art streaming anomaly detection method into an existing online learning framework. Given River’s focus on streaming and incremental learning, I believe MemStream is a strong conceptual and practical fit for the library.

Summary of the contribution

The implementation follows River’s coding standards, API conventions, and dependency constraints. It is based on the original code released by the authors of the MemStream paper and has been adapted to ensure full compatibility with River.

The contribution includes:

• A new module under river.anomaly introducing MemStream
• An abstract base class capturing the core MemStream pipeline
• Documentation and usage examples

Two concrete implementations:

• a denoising autoencoder–based MemStream
• a PCA-based MemStream

Documentation and usage examples

Before submitting or finalizing a pull request, I would greatly appreciate your feedback on whether this contribution would be suitable for inclusion in River, on the proposed placement of the implementation within the river.anomaly module, and on any modifications that may be required to better align the implementation with River’s conventions. I would be very happy to revise the contribution in response to your suggestions. Here is the link to the code: https://github.com/NicolasNoya/river/blob/feature/memStream/river/anomaly/memstream.py

Thank you for your time and for maintaining River.

Best regards,
NicolasNoya

P.S.: I have also included a notebook to illustrate and experiment with the model.
test_search.ipynb

[kulbachcedric]

Hi @NicolasNoya,
sorry for my late reply.
That would be a great contribution.
I have some remarks. You implemented your MemStream based on. numpy which would perfectly fit into river.
However, as one of you implementations is a deep learning approach (AE), I would put into discussion if your the PCA based MemStream approach would be a good contribution for river and the denoising AE version would be a better fit for deep-river (using torch)?

Also feel free to make a PR, so that we can easily check what impact your changes would have and if tests etc. are running through.

Best
Cedric

[NicolasNoya]

Just made the PR, waiting for your feedback when you have time!

[NicolasNoya]

Hello Cedric, Thanks for your reply! Sure, we would be happy to push the code, and I totally understand that the autoencoder approach does not fit well within the river library. It was mainly developed as a proof of concept. For the contribution to river, we will focus on the PCA-based implementation and adapt the code to better match river’s design philosophy. Over the weekend, I’ll work on cleaning things up and will aim to submit a PR by the beginning of next week. Thanks again, and have a great day! Best regards, Nicolás Noya El jue, 22 ene 2026 a las 8:12, Cedric Kulbach ***@***.***>) escribió:

…

Hi @NicolasNoya <https://github.com/NicolasNoya>, sorry for my late reply. That would be a great contribution. I have some remarks. You implemented your MemStream based on. numpy which would perfectly fit into river. However, as one of you implementations is a deep learning approach (AE), I would put into discussion if your the PCA based MemStream approach would be a good contribution for river and the denoising AE version would be a better fit for deep-river <https://github.com/online-ml/deep-river> (using torch)? Also feel free to make a PR, so that we can easily check what impact your changes would have and if tests etc. are running through. Best Cedric — Reply to this email directly, view it on GitHub <#1740 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZPX5IINPJGKA35RACDTB7T4IBZ6TAVCNFSM6AAAAACQNGQ4AWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNJWG4ZTEMA> . You are receiving this because you were mentioned.Message ID: ***@***.***>