Overview

Amazon has introduced COSMO, a commonsense knowledge‑graph framework that leverages large language models to infer the hidden reasoning behind product pairings and purchase intent. Launched in early 2024, the system moves search and recommendation logic away from pure keyword matching toward a deeper understanding of how shoppers think about products. Sellers who want their listings to surface in this new environment must adapt their content and keyword strategies accordingly.

Key Points

• Commonsense reasoning drives recommendations — COSMO interprets implicit relationships, such as recognizing that “slip‑resistant” is a key attribute when a shopper looks for “shoes for pregnant women,” even if the term never appears in the query.
• Two real‑world data streams feed the model — The framework trains on (1) query‑purchase pairs that capture what customers type and then buy, and (2) co‑purchase pairs that reveal items purchased together in the same session, after heavy filtering to discard accidental pairings.
• Benchmarking shows up to 60 % macro F1 gain — When evaluated against the KDD 2022 dataset, COSMO‑enhanced models outperformed baseline approaches by as much as 60 % in macro F1 score; fine‑tuned variants still delivered 28 % macro and 22 % micro F1 improvements.
• Human‑in‑the‑loop validation preserves relevance — LLM‑generated hypotheses are reviewed by annotators who discard relationships that are technically possible but unlikely in actual shopping behavior, ensuring the graph mirrors genuine consumer thought patterns.
• Conversational, layered search becomes possible — COSMO enables queries to evolve naturally—e.g., a shopper moves from “camping” to “camping air mattress” to “lakeside camping air mattress”—with each refinement surfacing products based on contextual understanding rather than exact keyword overlap.

How COSMO Works

1. Data collection — Amazon extracts two curated datasets from live sessions: (a) query‑purchase pairs that link a shopper’s search phrase to the final purchase, and (b) co‑purchase pairs that capture items bought together. The co‑purchase set undergoes aggressive filtering to eliminate coincidental combos such as shampoo paired with batteries.