msnews.github.io/assets/doc/introduction.md at master · msnews/msnews.github.io

Abstraction: MIND news recommendation dataset structure and collection methodology

Key points:

• Collected from anonymized Microsoft News behavior logs: 1 million users with 5+ clicks over 6 weeks (Oct 12 – Nov 22, 2019); MIND-small uses 50k users
• Four files per split: behaviors.tsv (impression logs and click histories), news.tsv (category, title, abstract, entities), entity_embedding.vec and relation_embedding.vec
• Entity/relation embeddings are 100-dimensional, learned from a WikiData subgraph using TransE; linked via WikidataId with type, confidence, and character offsets
• Impression log format: ImpressionID, UserID, Time, click history (ordered), impressions list with click labels (1/0); orders shuffled for impressions
• MIND-small contains only training and validation sets; full content bodies not available due to licensing; URLs have expired

Connections: Microsoft · Microsoft News · News Recommendation · Knowledge Graphs · Recommendation Systems

Source: https://github.com/msnews/msnews.github.io/blob/master/assets/doc/introduction.md