Docs for Multiclass Classification Metrics are wrong.

[antoniovs1029]

The docs on https://docs.microsoft.com/en-us/nimbusml/concepts/metrics#multiclass-classification-metrics are wrong since it states the following:

This corresponds to evaltype=’multiclass’.

[...]

(class N) - Accuracy of class N

Problem is that the (class N) columns that are the output of the evaluation aren't Per class Accuracies but actually Per class Log Loss metrics.

I've stepped into from ML.NET to NimbusML and this was the case.

For instance, ML.NET's MulticlassClassificationMetrics object doesn't hold a PerClassAccuracy member, but it does PerClassLogLoss

https://docs.microsoft.com/en-us/dotnet/api/microsoft.ml.data.multiclassclassificationmetrics?view=ml-dotnet

And in ML.NET's multiclass evaluator there's no code to calculate per class accuracy, only per class log loss: https://github.com/dotnet/machinelearning/blob/c255ac7c45af7a19a6e5ab1e26fdbe44874f241a/src/Microsoft.ML.Data/Evaluators/MulticlassClassificationEvaluator.cs#L298-L308

So I don't know if maybe in the past ML.NET/NimbusML actually output per class accuracies, but at least now it does seem the docs are wrong and would need to be fixed. Thanks.