I am comparing three methods decision tree, Naïve Bayes, and K-NN. I have a dataset with missing values and by using the Weka value replace tool I was able to replace the values.

I ran two test one before replacing the missing value and one after replacing the missing value. Before replacing the value k-nn correctly classified 55% instances but after replacing missing value it classified 56% correct instances.

I want wondering why this happens to k-nn but not the other methods of classification?

I doubt you'll get a good response on this forum but I'll start the ball rolling:

• If you only have small amounts of data (< few hundred) then small anomalies will look like large trends.
• Is your Weka tool putting in correct replacement values?
• Is it then correct to store these values in a database rather than just record that no value is known?
• Assuming the 3 methods all use different strategies to classify items then isn't it fair to guess that they'll use your new artificial data in different ways?
• Have you looked at your coursework notes?