ABSTRACT
Lack of student persistence and retention is significantly hurting the US in producing the required number of qualified graduates, especially in STEM fields. Although many factors contribute to students falling off track, one of the controllable factors is the identification of at-risk students followed by early intervention. Predicting the performance of students enables educators to single out struggling and highly talented students. Struggling students are often identified very late into an academic year, thus leaving little to no time for seeking consultation and determining the best course of action to improve performance. Some of such struggling students resort to dishonest means to catch up or make up at the last minute resulting in a higher number of academic integrity violations being observed and reported. Recently, the COVID-19 pandemic further corroborated the presence of such challenges. This research explores the possibility of using artificial intelligence to identify key elements in small datasets which could contribute to the development of a predictive student performance solution. A small set of data obtained through systematic data collection was used to train a predictive algorithm and aid in the analysis of in-class learning, which would lead to a viable student performance predictive solution. The data was collected for 133 students from a total of four sections of three different courses. With a limited amount of data, we were still able to construct a predictive solution able to produce valuable insights into the behaviors of students. The model's resulting accuracy on the test set is 0.85 and the model indicates that the earliest time to begin predictions is right after the midterm exam. The model performs well in its task to predict student performance and identify correlations between different variables. However, it is at this time subject to limited data which although treatable, can affect the accuracy and its ability to predict a final score numerically. This work paves the ground for future studies on the use of machine learning using in-class learning data, analyzing student learning as a function of time within each session rather than by grades alone.