Comment

Comment by Steffen Springer on October 27, 2009 at 3:18pm

Here is a paper which helped me to get a better feeling for the challenges in this area:
http://www.sigkdd.org/explorations/issues/4-2-2002-12/lo.pdf

Note: I studied this paper a year ago and did not do any reasearch since then. There may be some "fresher" work available

Comment by Richard Boire on October 26, 2009 at 1:22pm

Thanks Patrick for the info and I am looking at the various links that you have given me.

Comment by Patrick Surry on October 26, 2009 at 12:37pm

Oh, forgot to mention: there were a couple of good talks about uplift at the recent Predictive Analytics World event, including Eric Siegel’s keynote, Andrew Pole's Challenges of Incremental Sales Modeling in Direct Marketing, and Mike Grundhoefer’s case study from US Bank

Comment by Patrick Surry on October 26, 2009 at 12:34pm

[Disclaimer: I work for one of those vendors, Portrait Software, offering uplit solutions]

We've been working on uplift (aka net lift, incremental response, differential response) modeling for many years now. From my perspective, the most interesting feature of the problem is that you can't directly measure the incremental response of an individual. In a traditional response model you can make a prediction for probability of response, then measure the actual response (yes or no), and finally calculate an error for that individual. But in the net lift scenario you make a prediction of "change in response rate if targeted": since you can't both target and not target a single individual, there's no concept of a pointwise error estimate. This makes it non-trivial to even define the quality of an uplift model (and to compare two models) - there's no easy analogue to R^2 for example.

We've seen lots of approaches to building net lift models, but most are indirect in that they are based on estimating the response rate for targeted and untargeted customers and looking at the difference in predicted response rate if targeted vs if not. These models don't seem to work very well in practice: my intuition for why not is that they rely on the error terms from the two response models behaving nicely which isn't typically the case. Response rates are commonly ten times the incremental response rate, so the lift prediction gets completely washed away in the noise of the response predictions. We've been most successful with a direct modeling technique based on recursive segmentation.

There is some good discussion of uplift issues at http://scientificmarketer.com/search/label/uplift and of course you can find out more about our solutions at http://portraitsoftware.com/Products/portrait_uplift_optimizer