<img height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=1178143905890834&amp;ev=PageView&amp;noscript=1">

Uplift Modeling: The Math Behind Successful Promo Targeting

With the advent of machine learning, it is easy to build a model to predict whether a particular lead would convert by studying the data of different segments. But can this model be actually used to target promotional offers to the potential customers? Well, there is one critical glitch in this approach.

The model would surely tell us about the hot leads that can be converted. However, the same model does not tell us whether the lead will be, indeed, converted by the promotional offer. It is quite possible that he/she would have purchased it anyway! In order to study the attribution of a promotional offer campaign, we use Uplift Modelling.

Uplift models are used to identify customers who are most likely to respond positively as a result of receiving some kind of offer/discount. It focuses on the change in the target variable caused by marketing incentive (e.g. change in sales caused by a discount email) instead of focussing on the target variable itself (e.g. checkout).

 Let us try to understand the same using an example. Consider a promotional offer campaign in which customers have been sent offers/discounts, to increase their inclination towards purchasing a product.

We can consider a case of binary treatment (i.e. a customer may receive the offer or not) and a binary objective (i.e. a customer may purchase or not). The uplift in this case is the probability of purchasing (if they receive the offer) minus the probability of purchasing (even if they do not receive the offer). Since, the two events are mutually exclusive i.e. the same person cannot receive and not receive a particular offer, we try to build this model using a test and a control group.

An ideal control group is supposed to mimic the behaviour of the test group, in case all the external parameters are the same. The difference in this case would be, the test group is subject to an offer while the control group does not receive any offer. Since we have the required datasets, we can study the difference in the purchase behavior patterns of the 2 groups to understand the net incremental revenue (NIR) generated due to this offer campaign.

The above analogy enables us to divide the population into 4 segments, as shown below:

Uplift Modeling

  • Sure thing: Customers who would always buy the product, regardless of whether or not they receive any offer (0 uplift/NIR).
  • Persuadables: Customers who are likely to buy the product just because they received an offer (+ve uplift/NIR).
  • Do Not Disturb: Customers who became less likely to buy the product because they received an offer/communication (-ve uplift/NIR).
  • Lost cause: Customers who would never buy the product, irrespective of whether or not they receive an offer (0 uplift/NIR).

Uplift modeling estimates a causal effect of treatment and uses it to effectively target customers that are most likely to respond to a promotional campaign.*

Thus, the assertion behind uplift modeling is that we should target ‘Persuadables’ to send out our promotional offers, because they are the primary group which result in a positive impact from our campaign. Sending promotional offers to customers belonging to ‘Sure thing’ or ‘Lost cause’ would be a waste of our resources, and an opportunity cost if limited by the number of offers we can send. Sending out offers to ‘Do Not Disturb’ group would actually decrease sales.

If you want your promotional campaigns to use this modelling technique to improve your campaign RoI, reach out to beawesome@skellam.ai


About The Author

Leave a Reply