YSDA Recommender Systems Course

This repository contains materials for the Recommender Systems course taught at the Yandex School of Data Analysis. This branch corresponds to the ongoing 2026 spring semester.

Syllabus

• Week 1: Intro
  • Lecture: course overview and organizational details, intro to "Recommender Systems" problem
  • Practice: basic recommenders, user-item latent space
• Week 2: Candidate generation & metrics
  • Lecture: recsys metrics & candidate generation: classic ML, ANN, mixing
  • Practice: yambda contest overview & baseline solution
• Week 3: Ranking, diversity & metrics
  • Lecture: reranking - losses, algorithms, metrics; diversity control and MRR / DPP
  • Practice: classic algorithms (MF, SLIM, EASE); ranking - pool building, undersampling, composite targets
• Week 4: Deep learning for RecSys & neural candidate generation
  • Lecture: two-tower architecture, softmax model & sampled softmax loss, contrastive learning, negative sampling and LogQ
  • Practice: paper review on LogQ correction and negative sampling techniques
• Week 5: Neural candidate generation, pt.2
  • Lecture: cold start and long-tail, user / item encoding strategies, sequential models, beyond two-tower (GPU retrieval, generative retrieval)
  • Practice: aspects of training neural networks for RecSys
• Week 6: Neural ranking, pt.1
  • Lecture: why use DL, feature encoding (categorical, embedding, scalar), model composition
  • Practice: paper review on Piecewise Linear Encoding and Unified Embeddings
• Week 7: Neural ranking, pt.2
  • Lecture: feature interaction modelling, MLP, multi-task & knowledge distillation
  • Practice: paper review on DCNv2 architecture
• Week 8: System Design, pt. 1
  • Lecture: data architectures, logging, biases, data drift and monitoring
  • Practice: feature storages for different scales
• Week 9: System Design, pt. 2
  • Runtime design, candidate funnel, GPU inference, data delivery, controlled degradation, cold start
  • Practice: paper review on GPU retrieval - LiNR, SilverTorch
• Week 10: RecSys Transformers applications
  • Lecture: all about sequential models (transformers) on user action history
  • Practice: paper review on PinnerFormer & TransAct (by Pinterest)
• Week 11: Reinforcement Learning in RecSys
  • Lecture: all about bandits - algorithms, Thompson sampling, contextual bandits
  • Practice: applications of bandits & off-policy evaluation for e-grocery
• Week 12: Case Studies of Yandex's services
  • Lecture: recsys design for e-grocery, movie streaming platform, blogpost & news feed
  • Practice: deep dive in Lavka (Yandex's e-grocery) recommender system
• Week 13: Trends in RecSys
  • Lecture: state-of-the-art approaches & novel papers in RecSys
  • Practice: best contest colutions on yambda retrieval

Homeworks

• Week 1: minirec: basics and services - storages, simple components & service runtime
• Week 2: minirec: candidate generators - metrics, classic models (ALS, BPR, SLIM), vector indexes
• Week 3: minirec: ranking - candidate logging, training pool construction, model ranker, Mixigen
• Week 4: two-tower models - metrics, loss functions and negative sampling for candidate generation
• Week 7: neural ranking - feature encoding and a multi-task DCN V2 ranker with a mixture of experts
• Week 10: transformers for candidate generation - dataset, architecture, training and inference
• Week 13: Semantic IDs - bonus homework on generative retrieval

Staff

• Daniil Tkachenko
• Kirill Khrylchenko
• Roma Nigmatullin
• Alexey Krasilnikov
• Artem Matveev
• Vladimir Baikalov
• Vlad Tytskiy
• Ruslan Kuliev