DeepRecSys, лекция 2: ML дизайн рекомендательных систем

Lecture: Kirill Khrylchenko Seminar: Vladimir Baikalov This week we continue our journey through recommender systems from a more classical perspective and focus on the ML design of real-world recommender systems. The lecture is structured in a way that resembles a typical ML design interview for recommender systems. Imagine that you are stranded on an uninhabited island and need to build a recommender system from scratch. What would you do? We go step by step through the key components of such a system: 1. Metrics - defining goals and understanding what we optimize. 2. Data - logs, impressions, metadata, and their implications. 3. Retrieval - candidate generation, multi-stage design, and practical challenges. 4. Ranking - model inputs, objectives, and multi-signal optimization. 5. Bonus - discussion of the lecturer’s R&D experience. The seminar is complementary to the lecture: We review strong classical baselines that can be tried before deep learning. We analyze the Yambda paper, discuss baseline results, and highlight evaluation caveats.