From look-alike to uplift in customer lifecycle models

Vasily Sizov, Anastasia Grigorieva4 Jun 2022

Problem

A bank's small and medium business (SMB) team sells about 10 products through call centers, branches, email and SMS. Contacts are limited, so the bank must choose whom to contact about which product. Banking products mature slowly (up to 3 months), each has few target actions, and many have no campaign history. The team needed about 120 look-alike, response and uplift models and proof that they beat business rules.

Example from the article: a March 2022 call-center pilot (17 thousand customers per segment) for opening a current account (RKO).

Solution in brief

Build the three model types in order of data cost: look-alike from product-opening facts, response from contacted customers, uplift only where campaigns with a control group exist. Evaluate look-alike and response by coverage and balance ratio, uplift by the uplift curve, and confirm in a test/control pilot.

Questions this case answers
  • How to decide whom to contact with an offer when the purchase matures slowly and each product has few conversions
  • How to build and validate a targeting model on thousands of candidate features without leakage in cross-validation
  • How to choose between look-alike, response and uplift models for a bank's marketing campaigns?
  • How to build uplift models when there are about 2000 features and few target actions per product?

Tools and methods

  • Look-alike model — probability of the target action by similarity to past buyers.
  • Response model — probability of the action given a contact, trained on contacted customers only.
  • Uplift model — incremental effect of contact; two independent models and a single model with interactions.
  • scikit-uplift (sklift) — base for uplift meta-models, copied and extended inside the in-house framework.
  • CatBoost encoder, permutation importance, Boruta, SHAP — encoding, feature selection and post-analysis.
Continue reading — it's free

Google or a one-time email code. No password.