Evolution of Product Classification at Shopify: From Categories to Comprehensive Product Understanding

Summary

Shopify’s product classification system evolved from a 2018 logistic-regression model using TF-IDF into a product-understanding platform designed to capture categories, attributes, characteristics, metadata, and content-safety signals. A 2020 multimodal image-and-text approach improved classification, but growing product complexity led Shopify to combine Vision Language Models with its Standard Product Taxonomy, spanning more than 10,000 categories and over 1,000 attributes across more than 26 business verticals. The production pipeline uses two-stage category and attribute prediction, with category context shaping the second call, and runs on Kubernetes with NVIDIA GPUs; FP8 quantization, Nvidia Dynamo in-flight batching, and KV caching improve inference efficiency. A multi-LLM annotation system independently evaluates products and uses arbitration when models disagree, helping build training data for consistent predictions. The resulting system processes over 30 million predictions daily, delivers an 85% predicted-category acceptance rate, and doubles hierarchical precision and recall versus the earlier neural network approach.

Context

Shopify needed more granular product understanding than category classification alone, along with a consistent taxonomy, category-specific attributes, richer metadata, and content-safety and trust features as its platform and product diversity grew.

Approach / What changed

Shopify combines Vision Language Models with its Standard Product Taxonomy. A Dataflow pipeline makes separate category and attribute prediction calls, with the second prompt depending on the first result, while Kubernetes, NVIDIA GPUs, Nvidia Dynamo in-flight batching, FP8 quantization, and KV caching support production inference. A multi-LLM annotation and arbitration system supports training-data quality and consistency.

Takeaways

• The Shopify Product Taxonomy spans more than 26 business verticals, over 10,000 product categories, and more than 1,000 associated attributes, with hierarchical classification, standardized values, and cross-channel crosswalks.
• The prediction workflow treats category and attribute outputs as a coordinated process: both stages must succeed, partial failures trigger automatic retries, and outputs are validated against taxonomy rules.
• Shopify plans to migrate from a tree-based taxonomy to a Directed Acyclic Graph so products can have multiple valid categorization paths, better supporting cross-category products.

Related reading