attribute selection+weka+Naive Bayes

Question

I wonder which method among the following three methods is the best to perform an attribute selection:

1. using a meta-classifier
2. the filter approach
3. the native approach, using the attribute selection classes directly

The classifier that I'm using is Naive Bayes.

Could anyone guide me to find the best choice?

Answer 1

Well, there is no single answer.

1. You can use a decision tree classifier (like bagging) and select the attributes where the branching decisions are taken by the classifier. You can see the tree of course to see the branching and the attribute based on which the branching is made (and these attributes are important).

2. You can use forward selection or backward elimination technique. (a) in forward selection, use a single feature for which the error on validation/ test set is the least. Then with this feature included in your feature pool, try the rest of the features one at a time and choose the one that gives you, again, the least error. (b) in backward elimination, use all the features and take the error rate. Then eliminate every feature one at a time. Then select the one to take out of your feature pool that gives the maximum decrease in error.

Continue the process unless you are satisfied with your number of features (stopping criteria).

1. I personally use ranker algorithm and infogain attribute evaluator to rank the attributes first and then use either 2(a) or 2(b) to select attributes.

Errors- you can consider root mean squared error. Others can do good as well.

Answer 2

There is theory called No free lunch. You should try all three of them in your problem an measure results in your domain.