A unified deep learning framework for accurate student performance-prediction towards sustainable education

Abstract

Forecasting how students will perform early in their studies is difficult in higher education domain, reason being varying types of academic and behavioral data. Traditional deep learning (DL) techniques face challenges with the selection of influential attributes and feature relationships, resulting in poor performance prediction. To overcome this, our research presents a sequential attribute predictor (SAP) and an attention-enhanced hybrid deep learning (HDL) framework which combines feature-selection using SAP with an attention mechanism and examines five DL models: long short-term memory (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN), hybrid LSTM–GRU, and hybrid CNN–GRU. These models were trained on an extensive dataset collected from students of various universities in India and other countries. The assessment results showed that performance of hybrid models is superior to that of individual DL models. The best performing model was SAP based attention-enhanced hybrid CNN–GRU, which had lowest training MSE of 0.0434 and a validation MSE of 0.0284. A comparatively higher coefficient of determination (R²=0.9746) of proposed model clearly reflects a better fit between predicted and actual outcomes due to CNN’s capability to extract discriminative features, GRU’s effectiveness in modeling sequential patterns, and the attention mechanism’s ability to emphasize influential attributes.