Ranking User Attributes for Fast Candidate Selection in Recommendation Systems
Huichao Xue · 2020
Many recommendation systems use users' attributes to retrieve documents before ranking. Instead of using all attributes, this work explores algorithms that choose a subset, in order to achieve higher precision. We propose a model that forecasts the relevance of documents matched by each individual attribute. By restricting to top-K attributes based on the forecast, we observed 50% reduction in latency at 99th percentile on LinkedIn's job recommendation system, as well as increased users' engagements.