Amazon's COSMO Knowledge Graph Could Change How Products Get Recommended to Shoppers
Amazon introduced COSMO, an LLM‑driven knowledge‑graph framework that classifies product links into use‑case, capability, classification and causal categories. On the 2022 KDD product search benchmark it delivered a 60% macro F1 lift and kept a 28% macro F1 and 22% micro F1 edge after full fine‑tuning.
Overview
Amazon’s research team has introduced COSMO, a new framework that leverages large language models to construct commonsense knowledge graphs for product recommendations. Unlike traditional systems that rely solely on purchase‑correlation signals, COSMO attempts to infer the real‑world reasoning that links items—such as why a pregnant shopper might need slip‑resistant shoes. Sellers should watch this development because it could shift product visibility toward listings that articulate genuine use cases rather than just keyword matches.
Key Points
- Commonsense focus — COSMO moves beyond “customers who bought X also bought Y” to capture the logical motivations behind purchases, like safety needs for expectant mothers.
- Dual data streams — The model ingests both query‑purchase pairs (what shoppers typed and what they bought) and co‑purchase sessions (items bought together in a single order).
- Four relationship categories — COSMO classifies links into use cases, capabilities, classifications, and causal connections, enabling nuanced recommendation logic.
- Human‑in‑the‑loop refinement — Annotators review sampled relationships, extract guiding principles, and feed those back to improve subsequent model passes.
- Performance boost — On the 2022 KDD product search benchmark, COSMO achieved a 60 % macro F1 lift over frozen‑encoder baselines and retained a 28 % macro F1 and 22 % micro F1 advantage even after full fine‑tuning.
How COSMO Works
- Data ingestion — The system collects two behavioral sources: (a) the exact search query a shopper entered and the product they ultimately purchased, and (b) the set of items purchased together in one checkout session. For example, a query “water‑proof hiking boots” that results in a boot purchase, plus a simultaneous purchase of a trekking pole, feeds both streams.
- LLM‑driven inference — Large language models process the combined data to generate textual descriptions of product relationships across the four predefined categories. A model might output, “Slip‑resistant shoes are useful for pregnant women during kitchen work,” even if no product title contains both terms.
Analysis & Recommendations
Why This Matters
COSMO will favor listings that explicitly state functional contexts, target audiences and causal benefits, shifting traffic toward products with richer, scenario‑based copy. Sellers who don’t adapt may see reduced placement in "customers also bought" and "frequently bought together" sections.
Key Takeaways
- COSMO ingests two behavioral streams: query‑purchase pairs and co‑purchase sessions to infer relationships.
- The model classifies edges into four categories—use cases, capabilities, classifications, causal connections.
- Human annotators refine generated links, extracting principles that guide subsequent LLM inference passes.
- On the 2022 KDD benchmark COSMO achieved a 60% macro F1 lift over frozen‑encoder baselines and retained a 28% macro F1 and 22% micro F1 advantage a...
Recommended Actions
- →Edit product listings in Seller Central > Inventory > Manage Inventory to add explicit use‑case bullet points (e.g., "Ideal for chefs on slippery f...
- →Include audience descriptors and causal benefit statements in the description via Seller Central > Edit product details (e.g., "Designed for expect...
- →Track changes in recommendation placement through Seller Central > Advertising > Brand Analytics > "Frequently Bought Together" reports.
Comments
Join the discussion
Log in or create an account to share your thoughts on this update.
No comments yet. Be the first to share your thoughts!