Evaluating model - Lift and TPs

[Doug]

55% of the records in my dataset belong to the target class. I split up the dataset into 67% train and 33% test. I am now in the evaluation phase. In the test set, 74% of my target predictions are true positives and my lift in the 1st and 2nd deciles are 1.4 and 1.37, respectively.

I believe that 74% accuracy for target predictions is decent, but is 1.4/1.37 lift too low? That seems very low to me.

In your experience, what lift values do you typically get in real world situations?

[editor]

Since your target class is 55% of the records, the maximum lift is 1/0.55 ~ 1.8, so a lift of 1.4 in the first decile is not bad. In general, you should not expect lift close to perfect in real-world problems - if you get a near perfect lift, usually you have some false predictor.

See my KDD-99 paper on typical lift in business problems
Estimating Campaign Benefits and Modeling Lift,
Gregory Piatetsky-Shapiro and Brij Masand, KDD-99 Conference

We come up with a rule of thumb Lift(P) ~ 1/sqrt(P), under some conditions
see
http://www.kdnuggets.com/gpspubs/kdd99-est-ben-lift.doc