Amazon Product Search Ranking Events with Query and Item Metadata

A combination of Amazon ESCI and ESCI-S datasets in JSON-line format, used for fine-tuning bi- and cross-encoder models in the Metarank project. Each row represents a single ranking event with query and pre-joined item metadata.

Use Cases

• Fine-tune bi-encoder models using the 'query' and item 'title' fields for semantic search ranking.
• Train cross-encoder models on ranking events to improve relevance scoring between 'query' and item metadata.
• Analyze the relationship between 'query' phrasing and the 'title' of returned items to understand search intent.

Strengths

• Dataset combines two established Amazon product search datasets, ESCI and ESCI-S.
• Data is pre-joined, providing query and item metadata in a single ranking event record.
• Formatted for direct use in training ranking models like those in the Metarank project.

Limitations

• Unknown total row count and dataset size limit assessment of statistical power.
• No column list or sample data beyond a partial example hinders schema understanding.
• Geographic and temporal coverage are unspecified, limiting analysis of trends or regional bias.

Provenance

Source

Amazon ESCI and ESCI-S datasets, aggregated by Metarank.

Collection Method

Combined and formatted into JSON-line records with pre-joined item metadata.

Freshness

Last updated on 2023-04-14.