Anastasia Nikulina, Wildberries & Russ (behavioral ML models; previously responsible for the ML brand and the competition platform)
A talk about one non-trivial approach: how to reduce the number of CTR models for advertised products on a marketplace without losing quality. The idea: instead of per-subcategory models, cluster products by behavioral similarity and train one model per cluster. Clustering is done with the U-SPEC algorithm (spectral clustering for big data) on product embeddings from BERT4Rec. Outline: what a CTR model and the business metric are, why clustering was chosen, the U-SPEC architecture, re-clustering of products without embeddings, A/B results, and where else the algorithm can be applied.
Google or a one-time email code. No password.
