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

Shopify’s product classification system has evolved from basic categorization to an AI-driven framework using Vision Language Models and the Shopify Product Taxonomy. It accurately classifies products, extracts attributes, and processes over 30 million predictions daily, improving search, discovery, trust, and efficiency for merchants and buyers.

Shopify logo
Shopify
•
Kshetrajna Raghavan
•9 min read•advanced•
--
•View Original
KubernetesLLaMAModal

Overview

The article discusses the evolution of product classification at Shopify, detailing the transition from basic categorization to a comprehensive understanding of products using Vision Language Models and a structured product taxonomy. It highlights the improvements in product discoverability and the benefits for both merchants and buyers.

What You'll Learn

1

How to implement a multi-modal approach for product classification

2

Why Vision Language Models enhance product understanding

3

How to optimize inference pipelines for better throughput

Prerequisites & Requirements

  • Understanding of machine learning concepts and product taxonomy
  • Familiarity with Vision Language Models and their applications(optional)

Key Questions Answered

How has Shopify's product classification evolved over the years?
Shopify's product classification has evolved from a basic categorization system using logistic regression and TF-IDF classifiers to a sophisticated model leveraging Vision Language Models and a comprehensive product taxonomy. This evolution allows for better product understanding and improved search and recommendation experiences.
What are the key components of Shopify's current product understanding system?
The current system is built on Shopify's Standard Product Taxonomy, which includes over 10,000 product categories and 1,000 attributes, and Vision Language Models that provide multi-modal understanding, zero-shot learning, and natural language reasoning capabilities.
What optimizations are used in Shopify's inference stack?
Shopify's inference stack employs several optimizations, including FP8 quantization for reduced memory footprint, in-flight batching for dynamic request handling, and a Key-Value cache system to improve inference speed and efficiency.
What impact has the new system had on product predictions?
The new system processes over 30 million predictions daily, achieving an 85% acceptance rate for predicted categories, which enhances product discoverability and improves search relevance for buyers.

Key Statistics & Figures

Daily product predictions processed
30 million
This figure illustrates the scale at which Shopify's product understanding system operates.
Acceptance rate of predicted categories
85%
This high acceptance rate indicates the reliability and accuracy of the classification system.
Doubling of hierarchical precision and recall
2x
This improvement reflects the effectiveness of the new system compared to earlier neural network approaches.

Technologies & Tools

Some links below are affiliate links. We may earn a commission if you make a purchase.

AI/ML
Vision Language Models
Used for enhancing product understanding through multi-modal learning and natural language reasoning.
Data Structure
Shopify Product Taxonomy
Provides a structured framework for categorizing products and their attributes.
Infrastructure
Nvidia Dynamo
Facilitates model serving and improves throughput through dynamic request handling.
Infrastructure
Kubernetes
Used for orchestrating the deployment of the product understanding services.

Key Actionable Insights

1
Implementing a multi-modal approach can significantly improve product classification accuracy.
By combining image and text data, businesses can better resolve ambiguities in product descriptions, leading to more accurate categorization and enhanced user experiences.
2
Utilizing Vision Language Models can provide deeper insights into product attributes.
These models allow for zero-shot learning and natural language reasoning, enabling the classification of previously unseen products and generating rich metadata that enhances product listings.
3
Optimizing inference pipelines can lead to substantial performance improvements.
Techniques like in-flight batching and FP8 quantization help maximize resource utilization and maintain high prediction accuracy, which is crucial for handling large-scale product data.

Common Pitfalls

1
Relying solely on traditional machine learning methods for product classification can lead to limitations in handling complex product data.
As product diversity increases, basic classification systems may struggle to provide accurate results, necessitating the adoption of more advanced techniques like Vision Language Models.

Related Concepts

Machine Learning
Product Taxonomy
AI/ML Applications
Data Processing Pipelines