Problem
Instacart Ads shows sponsored products on several browse surfaces: Buy It Again, Frequently Bought With, Store Root, Collections and Item Details. Each surface has its own context and features, so the team trained a separate XGBoost click-through rate (CTR) model for every new surface. Click data stayed siloed, so a click on one surface did not inform the others. XGBoost could not take collection names, product IDs or search terms, because one-hot encoding them gives an impossibly wide input. Operations effort grew linearly with the number of surfaces.
Example from the article: a click on a product on Buy It Again never reached the other surfaces' models.
Solution in brief
Pool the training data of all surfaces and train one deep learning pCTR (predicted CTR) model with a Wide-and-Deep architecture. High-cardinality IDs and text go through embeddings, surface type goes to the wide side, and a factorization machine adds second-order interactions. Context features a surface lacks get default values, which the model learns to read as "absent".
Questions this case answers
- How to merge several context-specific click-probability models into one model when each context has its own feature set and the legacy models cannot take high-cardinality categorical inputs
- How to replace a growing set of per-surface CTR models with one model without losing quality on any surface?
- How to use high-cardinality features (user ID, product ID, collection name, search term) in a CTR model that XGBoost cannot take?
- How to share a user's click signal across product surfaces in ad ranking?
Tools and methods
- Wide-and-Deep network — embeddings of IDs and text (deep side); surface type and continuous features (wide side).
- Factorization machine layer — second-order feature interactions that vary by surface type.
- Mean target encoding (MTE) features — historical CTR per user-product pair and per context segment.
- XGBoost — the legacy per-surface baseline.
Google or a one-time email code. No password.