按年被提到
20222024 · 5 次2026
机器学习与深度学习›机器学习实践与反思·arXiv:2209.07663 · cs.IR

Monolith: Real Time Recommendation System With Collisionless Embedding Table

Monolith:带无冲突嵌入表的实时推荐系统
Zhuoran Liu, Leqi Zou, Xuan Zou, Caihua Wang, Biao Zhang, Da Tang, Bolin Zhu, Yijie Zhu, Peng Wu, Ke Wang, Youlong Cheng · 2022

字节跳动的实时推荐系统:无冲突嵌入表加在线训练架构

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 15 人首次 2022-11-06 · 最近 2026-02-20
15不同的人
9活跃月份
15讨论串
54被引(OpenAlex)

Monolith 是面向在线训练的推荐系统,指出通用框架在动态稀疏特征和训练与服务分离上不适合推荐场景。贡献包括:带过期嵌入与频率过滤的无冲突嵌入表以降低内存占用;高容错的生产级在线训练架构;并证明可以用系统可靠性换取实时学习。该系统已用于 BytePlus Recommend 产品。

HN 为什么提它

HN 上的人把它当作 TikTok 推荐系统公开技术细节的例子,称值得推荐系统爱好者一读;也有评论认为论文夸大了在线训练的重要性。

方法难度 3 / 5约 35 分钟经得起时间
读之前最好知道
1推荐系统基础
2嵌入表(embedding table)与哈希冲突
3分布式训练与参数服务器
4在线学习(online learning)

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
推荐CharlieSweeting · 2022-12

“they have made some genuine algorithmic/training improvements which are interesting.”

推荐它,认为 TikTok 推荐系统在算法和训练上有真实有趣的改进
TikTok’s Secret Sauce原评论 ↗
推荐Grimblewald · 2025-01

“It's an interesting read if you're into recommender systems or AI in general.”

推荐它,对推荐系统或 AI 感兴趣的人值得一读
TikTok preparing for U.S. shut-off on Sunday原评论 ↗
批评jamesblonde · 2026-02

“The Monolith paper is misleading that the 'online training' is key. It is not.”

TikTok's 'Addictive Design' Found to Be Illegal in Europe原评论 ↗
引用jerojero · 2023-11

“I mean, even Tiktok has papers detailing their recommendation algorithm [1].”

拿它说明连 TikTok 都发表论文详述其推荐算法
X is changing its algorithm to highlight smaller accounts原评论 ↗
引用wyldberry · 2024-12

“When you control the algorithm, everyone downstream of it is at the mercy of the power structures and incentives of your top level leadership [0].”

引它佐证掌控算法的人让下游用户受制于顶层权力与激励
TikTok divestment law upheld by federal appeals court原评论 ↗
引用supo · 2025-05

“If this was a solved problem, every e-com website would already run some variant of Monolith for all their shopping surfaces”

拿它说明若这是已解决的问题,电商网站早就都用上类似方案了
You Don't Need Re-Ranking: Understanding the Superlinked Vector Layer原评论 ↗
被提交为 story · 9 次