Use of SHAP values for identifying differences in behaviors for subpopulations under intervention

Files
Citation
Share
Date
Abstract
The advent of Artificial Intelligence (AI) is currently leading a new industrial revolution on almost all aspects of human life. Adoption of AI in traditional education has been lower than expected due to several reasons, including a lack of understanding of the processes behind it, which is fatal for situations like student dropout. An ideal AI tool for this problem would provide individually tailored interventions towards student retention, but that would require a much deeper understanding of what entails a successful intervention. Using a novel methodology for feature comparison between subpopulations, we found that the features obtained through our machine learning models coincide with both the opinion of interviewed mentors/tutors and with independently performed research with the same dataset origin, that the explanations obtained regarding student dropout match the real-world experiences of mentors and tutors, especially when dealing with highly explanatory features like previous average grades and interventions, and that additional beneficial features would be psychological and emotional well-being information. The results from our proposed methodology were validated directly by practicing mentors and tutors that deal with student dropout on a regular basis.