Recommender System Fragment 1

最新推荐文章于 2023-11-17 11:45:14 发布

原创最新推荐文章于 2023-11-17 11:45:14 发布 · 298 阅读

CC 4.0 BY-SA版权

1 篇文章

订阅专栏

本文详细介绍了推荐系统的处理流程，从获取用户特征开始，通过不同队列动态更新候选物品，利用预训练模型进行排序，并最终输出顶部推荐结果。文章还探讨了如何实现反垃圾及作弊模块等问题。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Recommender System Fragment_1

Procedure
- when a scenario(user) comes, fetch all the user’s properties and then pull candidates(items) from several queues depending on recall strategy(how? what model?). rank over all these candidate(~1000?) and push out the top ones.
Details
- each queue is updated dynamically.
- queue length? ~ 1000?
- always fetch from queue front? a bit slower and slower for bloom filter but acceptable cz queue not too long.
- ranking is a pre-trained stable model. just do ‘prediction’. parallel if needed.
- all queue and BF could be in memory.
Qs
- how to get user’s features? kv db and/or what? – kv db
- how to get doc features? are they stored in queue? – yes
- how and where to apply anti-spam & anti-cheating module?