Problem
Target runs a loyalty program, Target Circle, that sends members spend-threshold offers with a reward. The offer team had to raise engagement and incremental revenue by moving from cohort-level offers to a choice for each guest. Each guest has touched only a few offers, so the guest-by-offer matrix is mostly empty, and an offer's reward is unknown until guests respond.
Example from the article: "Make 2 qualifying purchases of $80 or more to earn a $15 reward in Target Circle earnings". The engine picks which such offer, with which amounts, each member sees.
Solution in brief
Non-negative matrix factorization fills the sparse matrix. Its output gives the reward signal and the offer features. A contextual bandit with a neural epsilon-greedy agent picks the offer for each guest. Seven A/B tests compared it with the production baseline.
Questions this case answers
- How to choose one item from a small set of candidates for each user when per-user feedback is sparse and the reward distribution is unknown
- How to evaluate a personalization policy online against a baseline while keeping a holdout for incremental impact
- How to pick the best spend-threshold offer for each loyalty member when most members have interacted with only a few offers?
- How to train a contextual bandit when the guest-offer interaction matrix is very sparse?
Tools and methods
- Non-negative matrix factorization (NNMF) — fills the sparse matrix, gives latent offer features.
- Contextual multi-armed bandit — offers are arms, guest features are context.
- Neural epsilon-greedy agent — three networks predict each offer's reward.
- Stratified A/B test with a holdout group — compares with the baseline, measures incremental sales.
- TensorFlow, PySpark, Kubeflow — models, data preparation, orchestration.
Google or a one-time email code. No password.