Insight into the Capabilities of Machine Learning Explainability Software Through the Evaluation of Two Prominent Open-Source Tools


Creative Commons License

Colorado V., Dubois P. A., Juma M., Topsakal O., Akıncı T. Ç.

4 th International Conference on Engineering and Applied Natural Sciences, Konya, Turkey, 20 - 21 November 2023, pp.46-58

  • Publication Type: Conference Paper / Full Text
  • City: Konya
  • Country: Turkey
  • Page Numbers: pp.46-58
  • Istanbul Technical University Affiliated: Yes

Abstract

The utilization of machine learning, deep learning, data analytics, and natural language processing technologies are becoming ubiquitous, and soon they will be part of diverse applications in real-world business scenarios, such as bank lending, stock market decision-making, and traffic monitoring. As machine learning continues to influence our daily lives, it becomes crucial for data scientists to comprehend the reasoning behind their models' predictions. This article aims to assess and analyze the effectiveness of two open-source tools, Alibi and Explainer Dashboard, in providing explainability for different algorithms and their outcomes. By doing this, we provide a sample framework for comparison and insights into the capabilities of explainability software. In this study, two distinct datasets were utilized, one pertaining to vehicle information, specifically cars, and the other concerning breast cancer diagnostic data. Upon evaluating the XAI tools, it can be concluded that there is room for improvement in open-source explainability solutions. To facilitate a comprehensive evaluation, a benchmark chart was compiled, focusing on three key metrics: satisfaction, trust, and effectiveness. Multiple questions related to these metrics were rated on a scale of 1 to 5. The ratings for each tool were aggregated, and their mean values were calculated to determine an overall score for each explainability tool.