pecos

Why we included this project

PECOS is built for a problem most ML libraries don't handle well: predicting from an output space that can reach hundreds of millions of candidates. It focuses on extreme multi-label ranking, where the task is to surface a small set of relevant labels from an enormous pool, and it offers two complementary approaches. X-Linear uses recursive linear models with fast C++ inference, while XR-Transformer fine-tunes pretrained transformers on multi-resolution objectives. The hierarchical label-tree design keeps both training and inference tractable at scale, which matters for teams doing product search, document tagging, or recommendation-style ranking. A bundled HNSW approximate nearest neighbor module handles sparse and dense features, so you get a practical retrieval tool alongside the ranking models.