Problem
Target's offer team needs to know how many times a customer will shop in the next few weeks. Offers and their goals work only when they fit how the customer really shops. The baseline, a non-sequential fully connected network called Trip Count Predictor, over-predicted trips, so many goals were unrealistic. Training data was also lopsided: highly engaged customers, who drive outsized value, were a minority.
Example from the article: a guest who normally shops twice in a period should get offers that fit naturally into those two trips.
Solution in brief
TargetRun is a sequence model: a bidirectional LSTM with multi-head attention reads 26 two-week steps of history, merged with static customer features, and outputs a probability for each trip count. Importance weighting fixes the audience mix, and focal loss focuses training on hard cases. A live A/B test against the baseline showed higher opt-in, completion and redemption.
Questions this case answers
- How to forecast the number of repeat purchases per customer over a short horizon when customer segments are heavily imbalanced
- How to size a personalized incentive goal to a customer's expected activity so the goal is neither too easy nor unrealistic
- How to predict how many trips a customer will make in the next two or three weeks
- How to set offer goals that customers can actually reach, so they redeem more
Tools and methods
- Bidirectional LSTM (Bi-LSTM) — reads weekly history in both directions
- Multi-head attention — highlights the weeks that matter most
- Static customer features — engagement level and membership details
- Importance weighting — rebalances the training audience
- Focal loss — up-weights hard misses, down-weights easy cases
- Live A/B test with a no-offer holdout — measures the effect on campaigns
Google or a one-time email code. No password.