Interviews with Leaders

Add a Comment

3 Comments

Comment by jayesh Sen on June 15, 2008 at 6:49am

When your dependent variable is binary in nature (independent variables may or may not be binary) then we use logistic regression to see the relationship among the variables.( dependency check of the binary variable) and in classification problem we don't have that dependent-independent criteria. we just want to classify the data set,so that a newly entered data can find its respective class. then we can use decesion tree.

Comment by Daniel Kocis on May 24, 2008 at 11:45am

Good question and the answer is ....
If you want/need to include ALL prospect variables into the solution AND expect a linear relationship THEN Logistic

ELSE

If you seek an interaction ONLY effect AND do not care about constraining yourself to hierachical paths THEN Decision Trees.

Binary classification is so much more powerful then dealing with distributions. The idea is to first develop an OPTIMIZED BINNING ALGORITHM for prospect variables and RE-DICREETIZE the outcome by weighting with co-variant candidates. From this matrix of OPTIMIZED (1,0) recodes, I prefer a combinatorial approach with filters.

Point is NEVER use all the variable or all information within a single variables distribution (too much noise). Focus on GREATER than Average intervals. LESS than Average intervals are interesting too if they have large N. I have developed something I call "Combinatorial Composition" that CLEARLY identifies Clusters based upon SIZE and LIFT criteria. Enough...

Comment by Geert Bilcke on April 13, 2008 at 10:50am

What do you choose for a binary classification : logistic regression or decision trees ? Why ?