Problem
A team builds a recommender that picks a promo offer for each customer by the highest purchase probability. The model looks good offline, but an offer can hand a regular buyer a discount on what he already buys. Customers respond, but revenue does not grow, because the discount cuts natural spend. Accuracy of the response model says little about profit.
Example from the article: at a fast-food chain (QSR) the offers were coupons for particular dishes. The purchase probability model had Gini about 0.7, but the pilot showed a negative economic effect.
Solution in brief
Pick the offer by a target function built from two uplifts: in purchase probability and in spend. Each uplift is the difference between a model that sees the offer's parameters and a model of natural behavior without an offer. The natural purchase probability weights the two terms, so customers at risk of leaving get retention-oriented offers and loyal customers get spend-growth offers.
Questions this case answers
- How to pick the offer that maximizes revenue rather than purchase probability when some offers reduce a unit's natural spend
- How to build an uplift model when treatments have parameters such as discount size and category, not just a sent or not-sent flag
- How to choose the offer that maximizes revenue, not just purchase probability, in a recommender system
- How to build an uplift model when offers have parameters (discount size, category), not just a sent/not-sent flag
Tools and methods
- Uplift with offer parameters — two models per outcome: with the offer and without it
- Purchase probability model — probability of a purchase with the offer and naturally
- Spend model — spend amount with the offer and naturally
- Target function — weighted sum of the two uplifts, weight is the natural purchase probability
- Exploratory pilot — collects the training data for the models
Google or a one-time email code. No password.