BARS-Match Overview#

BARS-Match: An Open Benchmark for Candidate Item Matching

Recommender systems generally comprise two main stages, matching and ranking. As the first-stage task, candidate item matching is designed to efficiently retrieve hundreds of item candidates out of the entire item corpus. Representative methods of candidate item matching include collaborative filtering, two-tower models, autoencoder-based models, sequential models, graph-based models, etc. To drive research in this direction, the BARS project aims to build an open benchmark for candidate item matching, which consists of: