dotnet / machinelearning

ML.NET is an open source and cross-platform machine learning framework for .NET.
https://dot.net/ml
MIT License
9.05k stars 1.88k forks source link

Periodicity detection accuracy is very low #7289

Open tappanajmera opened 2 weeks ago

tappanajmera commented 2 weeks ago

System Information (please complete the following information):

Description This is probably a question and not a bug. We tested periodicity detection from ML .Net and compared against other approaches. The accuracy is very low. Image

Accuracy selected means "accuracy on those where it gave an answer". So ML.NET gives a periodicity only on 28% of cases, and on those 28% it's 85% accurate. Kusto gives an answer on 94% of cases, and is 78% accurate on that.

Is this expected behavior?

LittleLittleCloud commented 1 week ago

Which trainer are you using

amueller commented 1 week ago

We're using https://learn.microsoft.com/en-us/dotnet/api/microsoft.ml.timeseriescatalog.detectseasonality?view=ml-dotnet if that was the question.