Search, ads, recommendation, and LLM applications

简体中文

This domain covers industrial recommendation, search, advertising, and their applications with LLMs. New industrial papers need quantified online A/B evidence or full-traffic or production deployment evidence explicitly accepted by the user. Classic baselines are included only as named exceptions.

Subfields

Stage or direction	Scope	Representative methods
Retrieval	Two-tower models, graph retrieval, semantic IDs, generative retrieval	YouTube DNN, TIGER, OneRec, PinRec
Coarse and fine ranking	Feature interactions, sequential interests, multitask learning, long sequences	DIN, DIEN, RankMixer, HyFormer, HSTU
Reranking and blending	List value, multi-objective constraints, source quotas, long-term value	SORT-Gen, Memento, DeGRe, DRL-PUT
Content understanding	Text/image/video representations, content–collaborative alignment, cold start	MM-LLM, PinCLIP, MIM, PRECISE
Moderation and risk control	Relevance quality, data review, privacy, anomalies, policy constraints	CORE, ASARL, RAMP, Proximity Features
Advertising and business decisions	CVR, delayed feedback, bidding, value-aware generation	ESMM, TWICE, SWAG, UniVA

Browse and evaluate

• Methods index
• Browse by company
• Browse by topic
• Browse by month
• Paper lineage and gaps
• Shared evaluation protocol
• Adapter and experiment overview

Moderation and risk control currently focus on relevance, training-data quality, privacy-constrained features, and policy constraints. This does not imply coverage of content safety, fraud detection, or advertising compliance. New work in these areas still requires real production evidence and should also report precision/recall, false positives, and business guardrails.