AutoKnow: Self-Driving Knowledge Collection for Products of Thousands of Types (Amazon, 2020)
Overview
This entry indexes an industrial publication from Amazon Science (2020): *AutoKnow: self-driving knowledge collection for products of thousands of types*. The original page is available at:
What the Work Is About
AutoKnow addresses the problem of building product knowledge at Amazon's scale — thousands of distinct product types, each with its own category structure, attribute schema, and attribute values. Manual curation does not scale, so the system is designed to be largely self-driving:
- Taxonomy induction: automatically discovering the category (taxonomy) structure relevant to each product type.
- Attribute discovery: determining which attributes (e.g., brand, color, material) are meaningful for each type.
- Attribute-value extraction: populating attribute values from noisy, large-scale product title and description data.
- It demonstrates how structured knowledge (taxonomy + attributes + values) can be mined automatically from unstructured catalog text.
- It highlights production constraints — data heterogeneity, long-tail product types, and the need for production-grade accuracy — that academic benchmarks often abstract away.
- The resulting product knowledge graph is a natural grounding source for retrieval-augmented and graph-RAG style systems.
- A Survey of Graph Retrieval-Augmented Generation for Customized LLMs
- RAG vs. GraphRAG: A Systematic Evaluation and Key Insights
- Algolia's Knowledge Graphs and Ontologies — Adding Knowledge to Keyword Search
The output of the pipeline feeds a Product Knowledge Graph, which is used to enrich Amazon's search, browsing, and recommendation experiences.
Why It Matters for IR / RAG Readers
AutoKnow is a representative example of industrial knowledge graph construction at e-commerce scale:
Note on This Post
The original forum entry for this item contained mostly template/placeholder content rather than a full translation. For exact experimental numbers, architecture details, and authorship, please consult the original Amazon Science page and the associated KDD 2020 applied-track paper. Do not cite quantitative claims from this summary without verifying against the primary source.