Hello, I have a data set with ~80 records, with ~8 features. I want to predict one of the features in future records. The feature is numeric and discrete. It ranges between -30 up to 140 with steps of 5. This feature doesn't normally distributed Until now I wanted to predict another feature which is boolean, so I used logistic regression. Which method should I use here? Maybe some kind of particle filter?

Thanks!

asked Nov 26 '11 at 18:37

Noam%20Peled's gravatar image

Noam Peled
16113

edited Nov 26 '11 at 18:45

2

If the feature you want to predict is numeric why not just use a linear regression and then round off the predictions to the closest discrete value?

(Dec 15 '11 at 11:01) Yevgeny

Thanks! This is kind of what I'm doing, but I have one problem: I know my data is noisy, so I don't want to predict an exact value. I prefer a tool which will predict how the feature will distribute given the other features.

(Dec 15 '11 at 11:08) Noam Peled

If the output variable is discrete, then what you went is classification not regression!

(Dec 17 '11 at 19:06) Noldorin

Indeed I can use classification, but I have more than 30 classes. Any thoughts?

(Dec 18 '11 at 10:43) Noam Peled

One Answer:

Have you considered matrix completion methods? The general idea is you've a matrix with some missing values (in your case features) and you want to impute these from the existing values. There are supervised and semi-supervised matrix completion methods that might suit your purpose. Check out the references section (and code) of this upcoming NIPS paper:

http://www.cs.berkeley.edu/~ameet/dfc

answered Nov 26 '11 at 19:21

Delip%20Rao's gravatar image

Delip Rao
6502810

Thanks! I'll take a look

(Nov 27 '11 at 08:43) Noam Peled
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.