Embeddings! Embeddings everywhere!
summary
Allegro describes a production recommender for ~100M offers and 20M accounts that learns item vectors with prod2vec using skip-gram, tunes hyperparameters to maximize MRR@25, and serves daily precomputed neighbors from a Faiss GPU Voronoi index (or per-category indices when too large). Content fallbacks use FastText and Inception V3 image features; orchestration uses Apache Beam and Luigi, storage in Cassandra. Experiments raise CTR almost 50% and revenue ~30% and proxy/metadata tricks enable recommendations for new items.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 6:05System design overview
- 8:12Product embeddings with prod2vec
- 13:27Content-based embeddings
- 14:36Nearest neighbor search
- 20:43Real-time serving
- 23:32Hybrid strategies
- 26:42User-to-item recommendations
- 30:16Hyperparameter tuning
- 32:00Data challenges
- 33:49Summary
- 34:45Q&A
related talks
28mEmbeddings
23mML-Powered Search & Recommendation 101
29mBuilding Generative Content Understanding Systems
27mData with Purpose: The Power of Embeddings
15mAdaEmbed: Adaptive Embedding for Large-Scale Recommendation Models
23mDeep Recommendations for e-Commerce
27mMachine Learning for Advanced AI Search
19mPOC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments