Overview

Amazon deployed COSMO, a framework using large language models to build knowledge graphs encoding commonsense relationships between products and customer intent. Rather than relying solely on purchase history and keyword matching, COSMO understands the reasoning behind customer choices, achieving up to 60% improvement in matching customer intent with relevant products across 18+ major categories.

Key Points

• Industry-scale knowledge graphs — Encodes millions of commonsense relationships mapping how customers think about products in terms of function, audience, location, and context
• Two primary data sources — Analyzes query-purchase pairs (searches followed by actual purchases) and co-purchase pairs (items bought in same session)
• Large language models generate hypotheses — AI analyzes behavioral data to infer implicit relationships, like inferring "men's shoes for wedding" means formal footwear
• 60% performance improvement — Production deployment shows up to 60% better matching of customer intent versus baseline models

How COSMO Ensures Quality

• Multi-stage pipeline — Employs intelligent preprocessing heuristics to filter noisy signals before AI generates knowledge assertions
• Critic classifiers validate output — Human-annotated training data filters hallucinated or low-quality assertions before entering production
• COSMO-LM specialized model — Custom fine-tuned language model for e-commerce commonsense reasoning produced millions of validated assertions from just 30,000 training examples

Seller Impact

• Optimize for use cases and context — Focus on clearly communicating product function, intended audience, and use-case scenarios in listings beyond just keywords
• Write natural descriptions — Well-written descriptions conveying how and why products serve customer needs help COSMO infer relevance without keyword stuffing
• Monitor co-purchase patterns — Products bought together influence how knowledge graph maps relationships, so ensure correct categorization and competitive positioning