Problem
Grocery stores carry hundreds of thousands of SKUs, so ordered items are often unavailable. A Dasher (delivery driver) shopping in the store must then call or text the customer to agree on a substitute, which is slow and tiring for both. A good substitute avoids a refund and a lost sale for the merchant. At launch there was almost no labeled data on which substitutions customers accept.
Example from the article: for a 12-pack of Coca-Cola, customers prefer a 12-pack of Pepsi or Dr. Pepper over a two-liter bottle of Coke, because quantity matters more than brand when buying in bulk.
Solution in brief
Customers choose acceptable substitutes in advance from model recommendations, so the Dasher needs no conversation. The model grew with the feedback: TF-IDF similarity with taxonomy rules, then a LightGBM classifier trained on thumbs-up and thumbs-down ratings, then a deep model on semantic item embeddings.
Questions this case answers
- How to recommend an acceptable replacement for an unavailable item in a large catalog when no labeled preference data exists at launch
- How to evolve a recommender from an unsupervised baseline to a supervised and then a deep model as feedback accumulates
- How to recommend substitutes for an out-of-stock item when there is almost no labeled data about good and bad substitutions?
- How to bootstrap a recommender in a new product vertical before user feedback exists?
Tools and methods
- TF-IDF cosine similarity on item names — unsupervised baseline for the cold start
- Catalog taxonomy heuristics — restrict candidates to relevant categories
- LightGBM binary classifier — probability that an item is a good substitute for the ordered item
- Thumbs-up / thumbs-down rating UI — source of labels
- Deep learning recommendation model (PyTorch) — item embeddings, dense features, explicit interactions
- Golden dataset and AUC — offline quality checks
- Online experiments — approval rate and coverage as input metrics
Google or a one-time email code. No password.