RecSys
The main goal is to alleviate item data sparsity and Cold Start problems. We address sparsity through Graph Recommender system, zero injection, or preference modeling. For cold start issues, we use Content-Based Filtering approaches.
Rather than focusing on SOTA, it's important to try various models optimized for the service using feedback from A/B testing. Overall, Collaborative Filtering like embedding-based approaches are outdated, with modern systems mainly focusing on Multi-stage recommendation with Candidate Generation and Ranking Model.
Input
Output is top-k recommendation
- Content Space - Content Latent Vector
- User Representation (session or user interaction history) - User Latent Vector
Scaling ability for the above two axes while maintaining minimal response time is important.
Search & Recommend System are in a same spectrum
Recommender Systems
Recommender System Notion
Recommender System Engines
Great overview
Tutorial
Pixel Screenshots is a feature that uses screenshots users take themselves instead of recording.
