Defaults for Classifiers/Regressors about jsat HOT 4 CLOSED

Classifiers/Regessors with reasonable default parameters already have (or should) constructors that fill in the options for you. Otherwise use the guess methods of the class to determine a reasonable distribution to search from.

from jsat.

Thanks for the rapid reply, Edward. I have no problem with using a default constructor for classifiers such as NaiveBayes() and NewGLMNET(), but this does not seem to be the case with Kernel-based classifiers. For instance, I am trying to construct a Forgetron, but it requires the parameters KernelTrick kernel and int budget. These options do not appear to fill in for me. I attempted to construct a KernelTrick object like KernelTrick kernel = new KernelTrick(), but was not able to instantiate the object in this way. Could you please provide insight into this situation?

from jsat.

As I said, those do not have no-arg constructors because there is no reasonable default value. Such things usually require a hyper-parameter search.

JSAT provides help to make that easier, with the Grid/Random search objects - and can auto-populate some parameters based on the aforementioned guess methods some classes have. See this blog post for some details, but if you don't understand Machine Learning well you need to be especially careful not to over-fit using these methods.

KernelTrick kernel = new KernelTrick(), but was not able to instantiate the object in this way. Could you please provide insight into this situation?

Its an interface, you can't instantiate interfaces - they define the contract.

from jsat.

Thanks for directing me to your hyperparameter example. By looking at the example, I figured out KernelTrick objects need to be declared as a parameter with the following syntax:
Classifier classifier = new Forgetron(new RBFKernel(),500);

I will refer back to the original papers to determine the parameters as well as via grid search. Thanks again.

from jsat.