Discretization

The front end.
Post Reply
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Discretization

Post by samoht »

Hello,

I have to discretize my spreadsheet data before i learn the parameters of my Bayes-Net with the EM-Algorithm. The net structure is already defined. The finer i discretize the data the more states for teh nodes i get.
Can I generally say the finer i discretize my data the better will be the result of learning the net parameters? Or can a too fine discretization have also a negative effect for the learning result?

Regards
samoht
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Re: Discretization

Post by mark »

No, because in the extreme you will get one state for each value in the data set, which is probably not what you what. You'll have to experiment and somehow evaluate what works best.
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Discretization

Post by samoht »

Hi Mark,

thanks for your reply. That means that if I discretize my data to extreme and then learn my bayes-net with this data that i will get very good results for this trining data. But if my data only slightly differ from my training data the net result can be very poor since for many bins there will be no or only few representative in the test data available.
So i have to make a trade off. Is this what you meant?

another Question: Why is the "bin count for original distribution" limited to 100? Is it not possible to have more than 100 bins?

samoht
Last edited by samoht on Thu Jan 06, 2011 9:32 pm, edited 1 time in total.
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Re: Discretization

Post by mark »

Yes, you will be overfitting the data. As for your other question, I am not sure, but it's probably what you describe it is.
Post Reply