Categorizing Products at Scale - Shopify

This post details the implementation of a product categorization system using Kesler's Construction to convert n one-vs-all classifiers into a single binary classifier. It leverages Logistic Regression and a greedy approach for inference on a hierarchical taxonomy. Text featurization uses HashingTF with PySpark, focusing on product title, description, collection, tags, vendor, and merchant-provided type. The system addresses scale and structure challenges of the Google Product Taxonomy (GPT) and is used by over 20 teams for features like marketing recommendations and market segmentation.