s1lver29 / recsys_course

The course on recommender systems conducted in National Research University - Higher School of Economics (Moscow, Russia). Academic year 2021-2022.

RecSys course

The course on recommender systems conducted in National Research University - Higher School of Economics (Moscow, Russia). Academic year 2022-2023 / Курс по рекомендательным системам, который проводится в Национальном исследовательском университете Высшей школе Экономики (Москва). Академический год 2022 - 2023.

Useful Links

• Wiki page of this course.
• The code materials for each seminars can be found in the corresponding folders /seminar*.
• To download any folder please use this link.
• Любые вопросы можно задавать в чат с технической поддержкой
• Table with grades

The most important section

The final grade is calculated as follows:

0.3 * Home Assignment + 0.15 * Article Summary + 0.15 * Weekly Quizzes + 0.4 * Exam

where Home Assignments - 3 home assignments in Jupyter Notebook (max 10 points each). Article Summary - конспект/презентация статьи из предложенного списка с критическим анализом (без выступления на семинаре) (max 10 баллов). Weekly Quizzes - 7 квизов по мотивам материалов семинаров, которые сдаются перед началом следующего занятия в Google Forms (ариф.среднее за все квизы, max 10 баллов за каждый). Exam - письменный экзамен в формате решения case-study построения рекомендательной системы для бизнеса (max 10 баллов).

Course Outline / Big plan for small victories

Week 1

Seminar 1

• Examples of RecSys models in production.
• Formalization of the ranking (recommender systems) task (2 popular types of tasks, 2 types of data sets).
• Ranking functions (BPR, WARP, RankNET, LambdaRank).
• Metrics for the quality estimation (Hitrate, Precision@k, Recall@k, MAP@k, NDCG@k).
• Taxonomy of RecSys approaches ([MF, FM, CF & other general], Content-based [including knowledge graph based, GB for ranking], Context-based, Sequential and session-based models, RL-based models, Hybrid [including two-level cascade]) approaches.
• Recommended sources on RecSys.

Seminar 2

• Hands-on example on the MovieLens dataset: movie recommender system.
• Basic baselines

Week 2

Seminar 3 - 4

• Item-based and user-based similarity, similarity metrics.
• Matrix Factorization (SVD et al.)
• Collaborative Filtering (ALS and iALS, HALS, NeuralCF)

Week 3

Seminar 5

• Content-based recommender models
• DSSM for RecSys
• Hybrid recommenders taxonomy
• LightFM (hybrid content model), Lightfm library

Seminar 6

• Gradient boosting for ranking task
• Example of cascade recommender model (using gradient boosting on the second level)
• Important preprocessing steps
• Cross-validation types

Week 4

Seminar 7

• Sequential models
• Next-basket and next-item prediction tasks

Seminar 8

• Context-aware recommender systems
• Time-aware и time-dependent models

Week 5

Seminars 9 - 10

• Autoencoders and Variational autoencoders for RecSys (VAE, Mult-VAE, Multi-VAE, Rec-VAE)

Week 6

Seminar 11

• Graph-based recommender systems overview
• Inductive learning (out-of-sample users, cold start problem)
• GNN, GCN, GraphSage -> PinSage, GAT

Week 7

Seminar 13

• Explainability & interpretability of recommender systems
• Knowledge-based graph recommenders

Week 8

Seminar 15 & Seminar 16

• Vanilla Production-ready RecSys service.

Contributors

• Marina Ananyeva
• Alex Milogradskiy
• Oleg Lashinin

License

All content created for this course is licensed under the MIT License. The materials are published in the public domain.