The course on recommender systems conducted in National Research University - Higher School of Economics (Moscow, Russia). Academic year 2023 / Курс по рекомендательным системам, который проводится в Национальном исследовательском университете Высшей школе Экономики (Москва). Академический год 2023.
- 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 final grade is calculated as follows:
0.3 * Home Assignment + 0.15 * Article Summary + 0.25 * Weekly Quizzes + 0.3 * Exam
where Home Assignments - 1 home assignments in Jupyter Notebook (max 10 points). Article Summary - конспект/презентация статьи из предложенного списка с критическим анализом (без выступления на семинаре) (max 10 баллов). Weekly Quizzes - 6 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).
Seminar 1-2
- 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
- Hands-on example on the MovieLens dataset: movie recommender system.
- Basic baselines
Seminar 3 - 4
- Item-based and user-based similarity, similarity metrics.
- Matrix Factorization (SVD et al.)
- Collaborative Filtering (ALS and iALS, HALS, NeuralCF)
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
Seminar 7
- Sequential models
- Next-basket and next-item prediction tasks
Seminar 8
- Context-aware recommender systems
- Time-aware и time-dependent models
Seminars 9 - 10
- Autoencoders and Variational autoencoders for RecSys (VAE, Mult-VAE, Multi-VAE, Rec-VAE)
Seminar 11
- Graph-based recommender systems overview
- Inductive learning (out-of-sample users, cold start problem)
- GNN, GCN, GraphSage -> PinSage, GAT
- Explainability & interpretability of recommender systems
- Knowledge-based graph recommenders
Seminar 15 & Seminar 16
- Vanilla Production-ready RecSys service.
All content created for this course is licensed under the MIT License. The materials are published in the public domain.