A gradient-boosted model estimates each user's chance of buying in November from their October behavior. It is useful because the probability it assigns is concentrated: a small slice of users accounts for a large share of the expected purchases.
Gain-based importance from the frozen model. Share of total gain per feature.
The frozen artifact, verbatim from metrics.json.
Predicted probability for every one of 75,915 pre-November users.
Co-purchase rules mined from same-session baskets. Pick an anchor to see what pairs with it.
20 products carry a rule. Pairs need at least three co-purchases and lift above 1.2.
A buyer of the anchor is 678.0 x as likely to also buy this as a random buyer.