Personalizing computer science education by leveraging multimodal learning analytics

This Research Full Paper presents an end-to-end framework to enhance personalized programming learning at a Higher Education Institution. A bundle of static and dynamic student data features (i.e. demographics, academic history and progression, clickstream or programming submissions) are collected to build predictive models. Artificial Intelligence and machine learning techniques are employed to automatically detect students-in-need and to adaptively provide them feedback in computer programming courses. At the macro-level, we hypothesize the bundled data features capture a meaningful set of predictors for each stage of the semester and each course. At the micro-level, being able to predict student's performances along with his/her progression provides automatic timely feedback. Three computer programming courses' historical student data was used to train the models. Predictions were then generated every week for the new cohorts of incoming students for those modules. A weekly report was sent to the lecturers with a rank of students and their associated probability of failing the next programming exam. The evaluation results show that, at the macro level, the predictions worked well with limited data. On one hand, student characteristics features are more relevant at the beginning of the course; on the other hand, the student's effort in the programming labs increasingly gain importance throughout the semester. In examining the micro-level assumption, we found that students who opted-in to receive personalized feedback indeed leveraged the generated predictions regarding to their predicted performance on their next exam. The personalized feedback includes various programming code recommendations, such as the programming code solutions from top-performers. We found that students corrected their programs after the programming suggestions from top-performers. Moreover, lower-performers improved their performance in the second laboratory examination with respect to students that were suggested programs but did not correct them and higher-performers respectively. In both scenarios, students who learned from the programs suggested and lower-performing students, showed a learning improvement over the two other groups. We should point out lower-performers had more room for improvement as they had lower grades in their first exam. Finally, we collected students' opinions about the feedback on the personalization intervention, to understand how it affects their behavior within the modules and if it encourages them to try new solutions or to revise material. Overall, it was very positive and most students would recommend this system to students attending the same course next year or would like to see this system included in other courses. Also, the last question captured their impression and how they would improve the platform. One student said: "It gave me confidence about the module. It gave me reassurance as to how I was getting on." while another said: "Good service, very helpful and effective way to manage your module". Higher-performers were getting an increasingly similar response were demanding a more personalized notification and some other additional learning resources.