Question

I am trying to use sklearn to predict a variable that represents rotation. Because of the unfortunate jump from -pi to pi at the extremes of rotation, I think a much better method would be to use a complex number as the target. That way an error from 1+0.01j to 1-0.01j is not as devastating.

I cannot find any documentation that describes whether sklearn supports complex numbers as targets to classifiers. In theory the distance metric should work just fine, so it should work for at least some regression algorithms.

Can anyone suggest how I can get a regression algorithm to operate with complex numbers as targets?

Answer 1

So far I discovered that most classifiers, like linear regressors, will automatically convert complex numbers to just the real part.

kNN and RadiusNN regressors, however, work well – since they do a weighted average of the neighbor labels and so handle complex numbers gracefully.

Using a multi-target classifier is another option, however I do not want to decouple the x and y directions since that may lead to unstable solutions as Colonel Panic mentions, when both results come out close to 0.

I will try other classifiers with complex targets and update the results here.