惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

AI
AI
Engineering at Meta
Engineering at Meta
T
The Blog of Author Tim Ferriss
Latest news
Latest news
Microsoft Azure Blog
Microsoft Azure Blog
D
Darknet – Hacking Tools, Hacker News & Cyber Security
Simon Willison's Weblog
Simon Willison's Weblog
M
MIT News - Artificial intelligence
V
Visual Studio Blog
N
Netflix TechBlog - Medium
P
Palo Alto Networks Blog
C
Cybersecurity and Infrastructure Security Agency CISA
阮一峰的网络日志
阮一峰的网络日志
P
Proofpoint News Feed
G
Google Developers Blog
MongoDB | Blog
MongoDB | Blog
V
Vulnerabilities – Threatpost
AWS News Blog
AWS News Blog
美团技术团队
博客园 - 聂微东
The GitHub Blog
The GitHub Blog
Stack Overflow Blog
Stack Overflow Blog
The Hacker News
The Hacker News
C
CXSECURITY Database RSS Feed - CXSecurity.com
L
Lohrmann on Cybersecurity
B
Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
爱范儿
爱范儿
Hacker News - Newest:
Hacker News - Newest: "LLM"
Hugging Face - Blog
Hugging Face - Blog
O
OpenAI News
W
WeLiveSecurity
Cisco Talos Blog
Cisco Talos Blog
Google Online Security Blog
Google Online Security Blog
T
Tenable Blog
Attack and Defense Labs
Attack and Defense Labs
C
Cisco Blogs
G
GRAHAM CLULEY
cs.CL updates on arXiv.org
cs.CL updates on arXiv.org
cs.AI updates on arXiv.org
cs.AI updates on arXiv.org
Y
Y Combinator Blog
Microsoft Security Blog
Microsoft Security Blog
Help Net Security
Help Net Security
The Last Watchdog
The Last Watchdog
S
Security @ Cisco Blogs
C
CERT Recently Published Vulnerability Notes
博客园 - 【当耐特】
T
Troy Hunt's Blog
Cloudbric
Cloudbric
IT之家
IT之家

博客园 - fariver

[PaperWritting] 多模态大模型架构图摘录 每日Paper - 2026-03-06 每日Paper | 2026年3月4日 [Paper Reading] Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking [PaperReading] OneSearch A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search [PaperReading] OneRec Technical Report [PaperReading] Generative Recommendation with Semantic IDs: A Practitioner’s Handbook [PaperReading] GME: Improving Universal Multimodal Retrieval by Multimodal LLMs [Paper Reading] UniME-V2: MLLM-as-a-Judge for Universal Multimodal Embedding Learning [PaperReading] UniME: Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs [PaperReading] Qwen2.5-VL Technical Report [PaperReading] DeepSeek-OCR: Contexts Optical Compression [PaperReading] SAIL-Embedding Technical Report: Omni-modal Embedding Foundation Model [PaperReading] VLM2Vec-V2: Advancing Multimodal Embedding for Videos, Images, and Visual Documents [PaperReading] VLM2VEC: TRAINING VISION-LANGUAGE MODELS FOR MASSIVE MULTIMODAL EMBEDDING TASKS [PaperReading] REACT: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS [PaperReading] MemGPT: Towards LLMs as Operating Systems [PaperReading] Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution [PaperReading] Mind Search: Mimicking Human Minds Elicits Deep AI Searcher [PaperReading] METAGPT: META PROGRAMMING FOR A MULTI-AGENT COLLABORATIVE FRAMEWORK
[Paper Reading] Tiger: Recommender Systems with Generative Retrieval
fariver · 2025-12-02 · via 博客园 - fariver

TIGER: Recommender Systems with Generative Retrieval

link
时间:NIPS 2023
单位:Google DeepMind
相关领域:Recommender Systems
被引次数:423

TL;DR

TIGER: Transformer Index for GEnerative Recommenders
现代推荐系统依赖大规模检索,本工作提出一种生成式检索的方法,即检索模型通过自回归方式解码出condidates的离散表示。具体而言,给定用户session(包含一个sematic ID序列),一个seq2seq的transformer模型将预预测下一个用户可能交互的item的sigmentic ID。
image

Method

Semantic ID的生成过程 以及 在生成式推荐中的用法 示意图

image

Semantic ID生成使用RQ-VAE算法

image

RQ-VAE Loss:
包含重建Loss与rqvae Loss
image
rqvae Loss包含两项:
第一项是让码本向量去靠近残差向量;
第二项是让残差微量去靠近码本向量;
image

Experiment

image

image

Q&A

Q:什么是Semantic ID Collisions?如何检测Collisions?如何修复Collisions?

  • 什么是?=> 在Semantic ID生成过程中,两个或多个不同的物品被映射到同一个Semantic ID元组的情况。当码本的尺度小于物品总数时,碰撞必然发生。即使组合数足够,相似物品也可能被量化到同一组码字。
  • 如何检测? => 检测碰撞的方法是维护一个查找表(Lookup Table),该表映射每个Semantic ID到其对应的物品。在为每个物品生成Semantic ID后,系统会遍历此表。如果发现一个Semantic ID对应了多个物品,则确认发生了碰撞。
  • 如何修复? => 修复碰撞的策略直接而有效:为发生碰撞的Semantic ID追加一个额外的、唯一的令牌(Token)。

Q:m级长度为K的residual的编码相对于一个m * K的编码有什么好处?
m级编码的好处是 层次化精细化,例如:

  • 第一级:捕捉物品最核心、最显著的语义特征(如“电子产品”)。
  • 第二级:在第一级的基础上,细化更具体的属性(如“手机”)。
  • 第三级及以后:进一步刻画细节(如品牌、型号等)。

Q:VAE的decoder还原的内容包括什么? => VAE的输入输出都是emb。

相关链接

https://zhuanlan.zhihu.com/p/1970625397411520943?share_code=4gjPSS7w8Nao&utm_psn=1979185503732508472