recsys-tutorial

ML meets Economics

Tutorials:

• implicit: full data & model pipeline, article
• LightFM: article
• How to build Item2Vec (or W2V) for item recommendations in retail

Classic RecSys:

• OK.ru: graph based recsys, article
• OK.ru: Neural item recommendations with cold start, article
• HH.ru: classic 2 level model of search at hh.ru, article
• Okko competition: classic 2 level model, article
• Yandex.Dzen: fit ALS -> fit Catboost on warm embeddings to predict warm&cold embeddings, 15-25min in video
• TikTok: No use of popularity features! post
• Instagram: Insights on candidate generation articles
• DoorDash: Store2Vec as a feature in recommendations
• Pinterest: Multi-taste user embeddings
• AirBnb: Hotel2Vec with novel positive samples approach

SOTA

• HRNN, Temporal-Contextual Recommendation in Real-Time

Search

• How to use W2V and FastText for search: Query2Vec
• Avito: FAISS for fast similar embedding search
• Similar vectors search with Nmslib (HNSW - hierarchical navigable small world), FAISS (embeddings space K-means clustering + Product quantizer) and Annoy (divides embeddings space with a binary tree)
• ElasticSearch basics
• DoorDash Elasticsearch meets logistic regression

Upsell

• Avito: Recommending additional item - upsell with advanced W2V

Uplift

• Directly optimizing Uplift in recsys by change in target

Optimization

• Avito Multi-armed bandits for item2item recommendations
• RecSys 2018 paper on multi-armed badits for explainable recommendations
• Pinterest: Personal notification volume optimization
• Uber.Eats: Multi objective optimisation in recsys, article
• Avito: Multi-objective optimization in search
• VK: Directly optimizing business metrics
• GlowByte Consulting: Customer communication chains optimization with RL

Cool articles:

• Measure user surprise by serendipidy metric