// the find
wangshusen/RecommenderSystem
A Chinese-language university course on industrial recommender systems, delivered as slides and lecture notes with accompanying YouTube/Bilibili videos. It covers the full production pipeline: retrieval (two-tower models, ItemCF, matrix factorization), ranking (MMoE, DIN, SIM), re-ranking for diversity (DPP, MMR), and cold start. Aimed at ML engineers who want to understand how large-scale recommendation systems actually work at companies like Alibaba or ByteDance, not how to build a toy movie recommender.
The coverage of industrial two-tower models is genuinely thorough — positive/negative sampling strategies, online serving constraints, and self-supervised extensions are all addressed separately, which most English-language resources skip. The cold start section is unusually practical, covering traffic throttling and cold-start-specific A/B testing rather than just the standard 'use content features' advice. MMoE and multi-objective ranking are explained with enough depth to actually implement them. The course maps closely to real production architectures, not academic benchmarks.
Everything is in Chinese — slides, notes, videos — which cuts off the majority of developers who'd benefit from it. There is no code: no notebooks, no reference implementations, nothing to run. The repo is essentially a structured link index to PDFs and videos hosted elsewhere, so if those links die the repo becomes useless. Last updated February 2024 with no indication of ongoing maintenance, and large gaps in the content (sections 8's reference paper is listed but several subsections from the README table of contents appear to have no slides).