lmxhappy/step-by-step-learning-ai

300+ paper notes on industrial recommendation, ads and search, across 16 topics, updated since October 2024. This is a notes repo, not a code or reproduction repo. Every note has a full English version.

🚪 Browse by topic (number of notes)

Behavior sequence modeling 54 · Generative recommendation 45 · LLM4Rec 42 · Deep retrieval 36 · Multi-task learning 20 · Semantic IDs 18 · Feature interaction 18 · Cross-domain 16 · Long tail 16 · Loss functions 15 · ML infra 12 · Representation quality 8 · Graph recommendation 7 · Item-to-item 5 · Feature selection 3 · Auto research 2

Plus 15 early paper lists (Oct–Nov 2024: paper titles with one-line tags, no write-ups).

Paper write-ups on industrial recommendation, ads and search systems: what the method does, what it gained in production A/B tests, and where the claims are weaker than they look.

Every note is written in Chinese for the WeChat account 稳扎稳打学AI and has a full English version next to it, with the suffix .full.en.md. The table below picks a few recent ones.

Recent highlights

Paper	Company	What it's about
TransX	LinkedIn	Compress the user's behavior sequence offline into 64 vectors stored as KV; online, candidates only attend to them. CTR +6.02% over an HSTU-style model.
BRIDGE	Kuaishou	Three decoder tricks for guessing Semantic IDs better: a codebook-anchored reasoning token before the first token, privileged EMA self-distillation on the student's own prefix, and a shared loop block that gives harder SID positions more passes. eCPM +0.65%, QPS +20%.
GR4AD-X	Kuaishou	Three parallel fixes to a deployed generative ad retriever: two first-level SID codes for boundary ads, serving-cheap capacity scaling, and per-prefix eCPM value heads in beam search plus a one-step rerank. Revenue +3.07%.
PAIR	LinkedIn	Embedding retrieval for Premium people recommendation: a platform-wide member embedding replaced by one fine-tuned on the surface's own clicks. CTR +7.5%, Premium retention +2.1%.
Pin-SCALE	Pinterest	Wiring Semantic IDs into two-tower retrieval for cold start. Fresh repins +3.67%.

Topics

All notes are organized by topic folder; open a folder and read the .full.en.md files. The full per-paper index is in the Chinese README.

Layer	Topics
Recommendation paradigm	Generative recommendation · LLM4Rec
Representation	Semantic IDs · Representation quality
Model architecture	Behavior sequence modeling · Feature interaction · Deep retrieval · Graph recommendation
Data & features	Feature selection
Training objective	Loss functions · Multi-task learning
Scenarios	Cross-domain · Item-to-item
Problems	Long tail
Engineering	ML infra · Auto research

Author: Mingxing Liu · WeChat lmxhappy · Zhihu. If this repo helps you, a ⭐ helps others find it.

This is a notes repo, not a code or reproduction repo.