Guild to Build a Great Rec System
I have worked at TikTok for two year and half in video recommendation system. TikTok has one of the best recommendation system in the world. I have learned a lot from my colleagues and I want to share my knowledge with you. I want to break down this series in 11 posts:
- Major Component in Rec System
- Modeling Target
- Training Data
- Loss Optimization
- Feature Engineering
- Hyperparameter Tuning
- Model Structure
- Reinforcement Learning in Rec System
- Federated Learning in Rec System
- AutoML and Application in Rec System
- Other
Today we are going to discuss our first topic: Major Component in Rec System.
Major Component in Rec System
In today’s industrial rec system, there are four major components: Recall, Rough Ranking, Fine Ranking, and Mix-Ranker. This design are used by most of the companies in recommendation and ranking system. Of course, some adjustment are made to fit the specific business needs. For example, TikTok user publish millions of videos and our FYP has to select 8 videos out of the pool in real time, thus computation cost and latency is a major concern. However, for some retail E-commerce such as DoorDash or Samsclub.com, the number of viable products are much smaller thus, thus the candidate pool is much smaller, so a rough sort might not be necessary. In this case, a recall + Fine will do the trick just fine.
In this post, I will give introduction to each component, and discuss some design concerns.
Recall
The goal for recall is to select a small subset of candidates from a large pool. This is usually achieved through Multi-Way Recall. Multi-Way Recall builds multiple recall models or rules to select candidates. Some commonly used model targets or rules are:
- Recall recent videos
- Recall popular videos
- Recall Regional Content
- Embedding Based Recall
- I2I: Item 2 Item
- U2I: User 2 Item
- Approximate Nearest Neighbor Recall
Reference
https://zhuanlan.zhihu.com/p/388603950
