Real-world uplift modelling with significance-based uplift trees [pdf]

Here's a plain language explanation of why uplift modeling is useful, written by the same author as the paper:

https://stochasticsolutions.com/uplift/

> It is normally assumed that the worst outcome direct marketing activity can have is to waste money. In fact, some direct marketing provably drives away business within certain segments, and it is not unknown for it to drive away more business in total than it generates. This is especially true in retention activity.

> [Non-Uplift] Churn and attrition models prioritize customers whose probability of leaving is highest. Such customers tend to be dissatisfied, so are usually hard to retain. To make matters worse, in many cases, the only thing currently keeping them is inertia, and interventions run a serious risk of back-firing, triggering the very defections they seek to avoid.

> It is more profitable to focus retention activity on those people who ... will leave without an intervention, but who can be persuaded to stay. Uplift models allow you to target them, and them alone. At all costs, you want to avoid targeting the ... so-called Sleeping Dogs, whose defection you are likely to trigger by your intervention. Again, uplift models can direct you away from those customers.

Interesting to see this paper here! Many years ago when working on a problem of offering online campaigns of some form, I had stumbled onto this paper, and it entirely changed my perspective. Eventually, I built an internal library based on the paper with a d3-based tree visualizer.

If memory serves right, my primary takeaway was that it isn't a good idea to make "customer retention" kind of offers to visitors (to a website) based on probability of purchase, because a fraction of them would have purchased irrespective of the discount offer. Of course,loyal customers should be rewarded in other ways, e.g., loyalty points, and this discussion is strictly for customer retention campaigns. In terms of model building this translates to: features that predict probability of purchase don't necessarily predict the difference in the probability of purchase for the same person given an offer. Of course, the second quantity is what we want. Its challenging to get to this since, in your data, for a given person, you would have made an offer to them or not - so you can't directly model this difference for them. Or model it in a statistically significant way. This paper provides a way to do so.

Great read. A little verbose for my taste. But lots of good ideas.