Backporting Clustering Indices to SciKit-Learn

[BradKML]

Currently I have asked this question for scikit-learn/scikit-learn#11369 (comment) and DistrictDataLabs/yellowbrick#1205 that there are some indices that are not in the major SciKit repos, and I see that your repo includes some of the included indices. However Online-ML is a specialized form of Machine Learning that is not necessarily useful for standard clustering, would it be possible to request a backport @hoanganhngo610 ?
https://github.com/online-ml/river/tree/main/river/metrics/cluster

[MaxHalford]

Hello there 👋

I'm sorry but I don't really understand your message. What are you suggesting exactly? What do you want to backport?

[MaxHalford]

I'm really sorry but I don't understand anything you're saying 😅. For instance, "update by the point" doesn't mean anything to me.

[BradKML]

@MaxHalford the Cluster Accuracy Metrics updates with x, y_pred, centers, sample_weight=1.0, which in the case of SciKit-Learn and others, are often Numpy Arrays of "number of points" X "dimensions of vectors" (with Labels being the other variable).
https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/metrics/cluster/_unsupervised.py

However, if that is the same case for River, the "total_points" would not make sense as the update method in the Cluster evaluation function.

• https://github.com/online-ml/river/blob/main/river/metrics/cluster/rmsstd.py#L67
• https://github.com/online-ml/river/blob/main/river/metrics/cluster/ssb.py#L71
• https://github.com/online-ml/river/blob/main/river/metrics/cluster/ssq_based.py#L73
• https://github.com/online-ml/river/blob/main/river/metrics/cluster/ssw.py#L254
• https://github.com/online-ml/river/blob/main/river/metrics/cluster/xiebeni.py#L79

[hoanganhngo610]

@BrandonKMLee As you may know, in online ML, we do not get to keep the information of all the points (and their labels) that have been passed through previously. This leads to the principle difference in the implementation of metrics of River and Scikit-learn, and as a work-around, we have to implement a counter in order to count the total number of points passed through in the data stream.

[BradKML]

Sorry for asking, but the folder for these indices are seemingly gone? I may have missed something. #484

[smastelini]

Hi @BrandonKMLee, these metrics are in the river-extra repo for the time being.