java - Weka Classifier Accuracy -

i've got 73,841 instances of data, 17 classes, using train classifier weka. data has been filtered using fft, , each instance has 3 points.

i.e. 85724.5409, 40953.2485, 3204935, 4539024.002345, ?/class

i've tried 3 classifiers: smo/j48/naive bayes.

the smo/naive bayes achieving accuracy rates of 16%

but j48 classifier producing accuracy rates of 98/99%.

questions:

can safely assume j48 classifier making sort of mistake? how can 2 results similar, , other different?
what can increase accuracy? there many classes, classes not separable?

thanks

i think output of decision tree inaccurate.

can provide tree generated or top 10 nodes see exact problem.

here of measures suggest improve accuracy.

class size: 17 classes indeed big. try reduce merging similar classes. (only done if doesnt affect scope of project.)
non-liner classifier: since having 17 classes linear classifier naive bayes/decision tree wont enough. did try non-linear svm or ensemble learning random forest. if enough data set present each of 17 class hmm choice better classification.

thanks, aravi

hj