I am interested in doing attribute selection using Chi Squared and the other AttributeEvals. This requires Ranker to be the search method. I know that the attributes can be pruned by selecting the best N or by providing a threshold. But is there a way to set a statistical significance level for the attribute like 95% confidence or 99% confidence level? I'm doing binary classification, and I want to be to specify that an attribute should have a difference in the two classes that is statistically significant.
↧