SMFSOP: A semantic-based modelling framework for student outcome prediction

Abstract

Over the past two decades, studying the various factors affecting student performance became essential. Knowing these factors assist in enhancing student’s performance, teaching practices and policy decisions. This research proposes a framework named “Semantic-based Modeling Framework for Student Outcome Prediction” (SMFSOP), to automatically map students’ activities within their learning environment to a standardized behavioral model (Community of Inquiry model (CoI)). The generated student representation is utilized to cluster students and predict an outcome based on their cluster. The framework is divided into three phases: Data gathering and pre-processing, automated mapping, clustering and prediction. The automatic mapping uses semantic similarity between student attribute names/descriptions, and CoI model indicators. Path and BERT similarities were identified as the best performers compared to human annotators. K-means, DBSCAN, and Kernel K-means are used for the clustering step, followed by LassoCV for regression-based prediction, & K-nearest neighbors for classification-based prediction. In order to prove that the proposed framework is generally applicable, three real life datasets were used as a case study. Best-performing trials enhanced outcome prediction as follows: In StudentLife Dataset, Adjusted R2 is enhanced by 3% (95% to 98%), and MSE decreased by 2.375 % (0.126 to 0.031). In social network dataset, Adjusted R2 was enhanced by 17% (65% to 82%). The MSE decreased by 4.4% (0.164 to 0.12). For the “Open university learning Analytics dataset” (OULAD), accuracy is improved by 1.56%, F1-score enhanced by 0.014. Precision is enhanced by 3.1%.