Replies to This Discussion

Permalink Reply by Neil McGuigan on April 26, 2011 at 11:47am

you could use principal components analysis

 

or, 

 

try variable selection techniques such as stepwise or genetic ones

• ▶ Reply

Permalink Reply by Ravirajan.K on April 28, 2011 at 5:31am

Thanks Neil. I will tyr.

• ▶ Reply

Permalink Reply by Mike Olson on April 27, 2011 at 4:51pm

I'd also recommend trying principal component analysis.

 

You could also use a decision tree algorithm (like C4.5) to generate a tree with limited depth, to figure out which variables are giving you the most information.  Then you can throw out the rest.

 

A third option would be to see if you have any variables that are highly correlated, and keep only one out of each set of correlated variables.

• ▶ Reply

Permalink Reply by Ravirajan.K on April 28, 2011 at 5:30am

Thanks Mike. I will try these methodologies from my end and keep you posted.

• ▶ Reply

Permalink Reply by Puneet Agarwal on April 28, 2011 at 10:13am

I would say Principal component analysis is a good way of reducing the number of variables but the PCAs would not make sense when you try to implement a logistic regression model or to decide the strategies. In my personal experience, It is very hard to explain PCA variables to the business partners and get them to digest the fact that each PCA is a combination of all the variables.

 

You can try stepwise logistic regression method, use VIF and reduce multicollinearity, check correlations between variables, check fill rates for variables and remove variables with less than 60% fill rate as you would need to impute the remaining observations. You could also try variable clustering methodology.

• ▶ Reply

Permalink Reply by Alex Zolot on May 26, 2011 at 10:40am

in R something like that:

 

install.packages('randomForest')

library(randomForest)  

 

rf = tuneRF(Xvars, Yvar, stepFactor=1.2, doBest=T)

rfi = rf$importance   

barplot(rfi[order(- rfi)])

 

• ▶ Reply