Model Apllier problem
Contributor I
Hi,
I created a model, learned on a training set with several nominal attributes (e.g. MONTH, OVERALL_STATE), everything seemed fine until I applied my model on a test set (containing only one example). I have following warnings in a log:
[Warning] KernelDistribution: The number of nominal values is not the same for training and application for attribute 'MONTH', training: 12, application: 1
[Warning] KernelDistribution: The number of nominal values is not the same for training and application for attribute 'OVERALL_STATE', training: 12, application: 1
When I put another 11 examples with range of my month nominal attribute, warning for month disappears. But I want to classify just one example, how should I do that without warnings? When I output values of DataRow for each attribute (model creation and application is done from Java code), all nominal attributes have value 0 (instead of values given in test set), what I am doing wrong? Thank you for your help.
Greetings,
dusankat
I created a model, learned on a training set with several nominal attributes (e.g. MONTH, OVERALL_STATE), everything seemed fine until I applied my model on a test set (containing only one example). I have following warnings in a log:
[Warning] KernelDistribution: The number of nominal values is not the same for training and application for attribute 'MONTH', training: 12, application: 1
[Warning] KernelDistribution: The number of nominal values is not the same for training and application for attribute 'OVERALL_STATE', training: 12, application: 1
When I put another 11 examples with range of my month nominal attribute, warning for month disappears. But I want to classify just one example, how should I do that without warnings? When I output values of DataRow for each attribute (model creation and application is done from Java code), all nominal attributes have value 0 (instead of values given in test set), what I am doing wrong? Thank you for your help.
Greetings,
dusankat
0
Answers
this is only a warning, not an error. It's only intention is to catch the users attention to control, if the datasets are really equal. So the nominal values that were present during training and now are gone are interpreted as an indicator, that the example set might not be of the same signature.
Greetings,
Sebastian