search

jeffheaton / encog-java-core

http://www.heatonresearch.com/encog
Other
742 stars 268 forks source link

Bayes training: wrong values in *_output.csv #73

Closed PetrToman closed 12 years ago

PetrToman commented 12 years ago
  1. In Workbench, create .ega file for Bayes classification - using data: http://dione.zcu.cz/~toman40/encog/data2.zip
  2. Double click .ega file and execute all tasks.
  3. Open data2_output.csv and check y vs. Output:y columns: Output:y contains 2 when there should be 1.
seemasingh commented 12 years ago

Not totally sure this is an issue. The data2.zip file did not train very well, so the output is going to contain many errors.

PetrToman commented 12 years ago

Here's a larger data sample: http://dione.zcu.cz/~toman40/encog/data7.zip

In any case, I don't see any reason why there should be values other than 0/1 - in case of a binary classification (when the ideal ouput is only 0/1).

I also suspect this issue causes that "Training error" calculation is wrong when the training finishes ("Training error" is currently not updated during the training in Workbench), because the ideal/real values do not (cannot) match.

jeffheaton commented 12 years ago

This was a result of the wizard classifying this as a regression problem, which cannot be solved by Bayes. Added in an error message for trying to force Bayes to do regression, also made the wizard use classification in this case, which it should have. Now gets better results.

PetrToman commented 12 years ago

It would be more consistent if Workbench displayed an error message when regression is selected - like for SOM: "Can't use a SOM with regresion."