when using filtered classifier we are getting different results than when
filtering during preprocessing and then doing classification. base
classifier RBF classifier and filtered classifier is the RBF network. We
think when using filtered classifier the error and classification attribute
that is generated during preprocessing are not passed to the base
classifier. How can we do this?
If you switch to evaluation on the training data in the Classify panel for both cases, you should get the same result in both cases.
In the Preprocess tab, the filter’s internal model (if there is one) will be build from the *entire* dataset. In the Classify panel, the filter model used for the textual output of the FilteredClassifier (e.g., a decision tree) will also be build from the *entire* dataset. However, if you use 10-fold cross-validation or a percent split evaluation, the filter model(s) in the FilteredClassifier(s) used for computing performance statistics will be build from the data in the training folds/split only, to prevent getting optimistic performance estimates.
In particular, applying a supervised filter in the Preprocess tab is dangerous and must be avoided when estimating predictive performance.
The class attribute is what is used as the independent variable by WEKA's Classifier implementations (e.g., the decision tree learners) and other schemes such as supervised filters and attribute selection methods.
If you perform tasks such as association rule mining or clustering, the "class attribute" will be treated just like any other attribute by default.
You might want to take a look at the online courses on using WEKA. They are free and currently open: