How Amazon Marketplace evaluates content quality with Amazon SageMaker AI
Learn how Amazon Marketplace evaluates A+ Content quality at scale on Amazon SageMaker AI by rendering each product page as a single snapshot, scoring it with a vision-language model fused with OCR readability signals in a CatBoost classifier, and using SHAP to turn each decision into prioritized, actionable recommendations for selling partners.
Amazon Marketplace utilizes Amazon SageMaker AI to evaluate the quality of content on A+ Content pages, which are rich media widgets on product detail pages. These pages feature visual assets that allow brands to communicate their unique value to customers, aiding them in making informed purchasing decisions. Brands, regardless of size, leverage A+ Content, but maintaining consistent quality across all content is challenging.
Traditional methods of evaluating each asset individually do not capture the holistic quality that emerges from how the various assets interact. Thus, the multi-model fusion approach, as described in this post, can help achieve this goal.
A+ Content offers brands creative freedom in expressing their product story, without strict content specifications like image ratios or text word counts. While this flexibility allows for compelling storytelling, it also leads to significant variation in content composition and quality. To address this issue, Amazon SageMaker AI employs a Vision-Language Model (VLM) to assess the complete content experience, providing actionable recommendations for improvement.
This approach combines the semantic scores from the VLM with readability signals from optical character recognition (OCR) using a CatBoost classifier trained on expert annotations. The resulting quality signals are then used to generate improved visual assets. The end-to-end pipeline for this process is implemented on Amazon SageMaker AI, encompassing data ingestion, annotation, model hosting, explainability, and production monitoring.
Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
