Analyze data sharing platforms

[PGijsbers]

We want to have a fresh look at how other platforms handle data sharing from a UX perspective.
This goes beyond Kaggle/HF, but should also have a look at how and where people in other sciences share their data, and why.
A strong focus on tabular data for now.

[SubhadityaMukherjee]

• HF : Folder of parquet files , with the occasional metadata csv
• Kaggle : Pretty much anything, but usually a folder with files and metadata
• NCBI (Bio tech info) : A tab delimited file : DataSet SOFT (but saved as a .gz archive)
• Fastai : tar.gz with all files in it and a metadata csv
• COCO dataset : good for checking image tasks - # Detection | DensePose | Keypoints | Stuff | Panoptic | Captions
  • folder with files and label splits
• U.S. Department of Energy Office of Scientific and Technical Information (Physics): Images are in .zip, tabular in .csv and metadata as .xml

[PGijsbers]

And beyond file formats, what is their upload process like? And what do you get after you upload? Easy citation? Good visualization? Great findability? Is it a requirement for conferences? And so on. I think we want a more holistic view of the process to help understand what drives people to those platforms.

[SubhadityaMukherjee]

I presented a small segment on this at the OpenML hackathon this month. Here are the relevant notes with more detailed information about this issue - link