Problem
Wildberries ranks advertised products by predicted click probability (pCTR), at about 600+ million events per day and 2+ million advertised products. A single global CTR model was not accurate enough on the business metric, and one model per parent category was still not enough. A model per subcategory worked, but there are more than 8000 subcategories, so training and maintaining that many models costs too much in people and infrastructure.
Example from the article: the 67 first-priority subcategories (160+ thousand products) were regrouped into at most 10 behavioral clusters, and a 7-day A/B test compared cluster models with subcategory models.
Solution in brief
Group products by behavioral similarity instead of by catalog category, and train one CTR model per cluster. Clustering runs on BERT4Rec item embeddings with U-SPEC, a spectral method that scales to millions of points. Products without behavioral embeddings are assigned to clusters by a small classifier on text and image embeddings.
Questions this case answers
- How to reduce the number of per-segment prediction models to a small set by grouping millions of items by behavior, where the grouping must scale to millions of points
- How to assign new items without behavioral history to behavioral groups using their descriptive features
- How to cut the number of per-category CTR models without losing prediction accuracy
- How to cluster millions of items on behavioral embeddings when K-Means gives unstable clusters and spectral clustering does not scale
Tools and methods
- U-SPEC (ultra-scalable spectral clustering) — clusters items on a sparse graph to representatives
- BERT4Rec item embeddings — behavioral features, 384 dimensions
- UMAP — dimensionality reduction in the comparison
- K-Means variants, HDBSCAN — baselines
- Relative Accuracy@20%, Subject Pass Rate — business metrics for choosing the method
- FC-block classifier — assigns items without behavioral embeddings to clusters
Google or a one-time email code. No password.
