Students' Performance Prediction Using Multimodal Machine Learning

Abstract

Technological intervention in the field of education has gained significant relevance, especially during the post-pandemic era. The three dimensions of interaction that influence learning are the student's interaction with the content, peers, and instructors. Learning ecosystems are expected to ensure these interactions in a seamless way. Technological interventions have provided us with provisions to establish the interactions. The data that we obtain while the student is interacting with content, peers, and instructors can serve as feedback to students and instructors. The motivation of the current study lies in the direction of investigating ‘what’ and ‘how’ current practices of establishing the interaction with content, peers, and instructor are influencing students' performance. The other dimension includes how demographic factors like gender influence the performance of students when technological interventions are made.The sample considered in the study included 140 first-year engineering students in a private university. The outcome of the study helped to do early prediction of student failures and identification of factors that influences the student's success. The data for the study was collected from multiple modalities. Clickstream data was collected from a learning management system to understand the interaction of students with the course content. Student collaboration data was collected from GitHub to understand the interaction of students with peers. Demographic data was collected from student academic performance to understand how past performance and demographic factors influence future performance.The findings reveal that the student interaction with the content and the student performance have a positive relationship with a correlation coefficient of 0.68. The algorithms including random forest, naive Bayes, decision tree, support-vector Machine, and extreme gradient boosting were used to perform multiclass classification to predict the performance. The students were grouped into four classes including ‘Excellent’, ‘Good’, ‘Average’, and ‘Poor’ using decision tree with a classification accuracy of 96%.