XAI-Scholar

Code and data for deriving empirical trends in XAI.

Sources

Please check the accompanying Medium blogpost and Arxiv report.

The data was collected as described in the above links. To recap, it is a mix of keyword search via SemanticScholar, and manual curation. The final collection has 5199 papers as of December 31st 2022.

The code used to interface with SemanticScholar uses the unofficial semanticscholar python library. Install with pip install semanticscholar. The plots and graphs use Seaborn and GraphOnline.

Format

XAI-Scholar_analysis.ipynb is a Jupyter Notebook with the code necessary to reproduce the results in the medium/arxiv reports.

xai-scholar.json is a json dictionary with 5199 SemanticScholar paperId keys.

Each paper has the standard fields given by the SemanticScholar API:

• paperId
• externalIds
• url
• title
• abstract
• venue
• year
• referenceCount
• citationCount
• influentialCitationCount
• isOpenAccess
• fieldsOfStudy
• s2FieldsOfStudy
• tldr
• publicationTypes
• publicationDate
• journal
• authors

Note: Some of the papers are missing some of the fields, or they are marked as empty or None.

The following fields are missing in xai-scholar.json and need to be retrieved from SemanticScholar due to their large size (expect around 1 GB without embedding):

• embedding
• citations
• references

The code to do so is in the jupyter notebook - but it simply calls semanticscholar.get_paper(paperId) for each paperId.