I like the new column with the ROC curve plots but for the other columns, I preferred the 2D plots instead of the 3D plots.

Thanks @MaiRajborirug, this is a nice visualization but I agree with @ogrisel: with the 3D it's harder to see the specificity of each of the estimators.

I particular, on the 2D plots it was easier to see the shape of the decision boundary with the black contour line.

Thank you for your reviews! I will create an ROC curve and accuracy score in the 2D-plot so that we have the performance comparison measurement.

According to your advice, I add ROC curves (last column), AUC, and prediction accuracy to the 2-D plots

The plot:

The last update is to make this PR a bit shorter.

The plot:

This is nice plot but I am a bit ambivalent about the usefulness of ROC curves for such toy examples. If we want to make an example with ROC curves @ogrisel suggested (more than 2 years ago) to change 2 of the benchmarks to an example. This would maybe be a better thing to do.

TBH, I had exactly the same reaction as @albertcthomas. I don't think that the quantitative analysis on such toy datasets is a must-have (maybe only the accuracy because it does not clutter the example so much). I think that the main point of the example is indeed a qualitative analysis. It provides highlights and intuition regarding the implemented algorithms, linked to assumptions made regarding the methods.

However, I agree with you that we miss an example where we should show an end-to-end pipeline where anomaly detection is beneficial in classification and this should be rigorously analyzed with such classification metrics/plots.

Another limitation of the ROC is that we only have 3 of the 4 methods as well.

@albertcthomas, I created a new PR #16606 corresponding to @ogrisel 's and your suggestion. Could you take a look at them?