Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) for Forecasting Learner Attrition: A Systematic Review

Abstract

The issue of learner attrition is a long-standing problem in Open and Distance Learning (ODL) settings where the lack of physical interaction and flexibility exacerbates the risk of disengagement. The use of Deep Learning (DL) techniques for forecasting complexity of behavioural and relational patterns of educational data has grown in usage. While Artificial Intelligence, DL in particular offers superior accuracy in forecasting attrition, the selection of appropriate techniques that addresses temporal sequence and relational patterns remains a critical gap due to inductive biases of ODL settings. This paper performs a systematic review based on the Preferred Reporting Items of Systematic Reviews and Meta-Analyses (PRISMA) tool in order to synthesize the current body of knowledge regarding the use of LSTM and GNN in forecasting attrition. The peer-reviewed articles were located in major digital databases and filtered based on predetermined inclusion and exclusion criteria. The review evaluated model archetypes, data properties, metrics of evaluation, and performance results. Results showed that LSTM models were more useful in learning temporal patterns of engagement, whereas GNN models were efficient at learning relational and social learning patterns. Nevertheless, differences in datasets, validation procedures and evaluation metrices made it difficult to directly compare the results. The study identified methodological gaps of single models and recommended the use of hybrid methods for increased accuracy. The review gave consolidated information that direct researchers and institutions in the selection of suitable hybrid deep learning model in forecasting learner attrition.