Early prediction of mid-term and final scores using deep learning models

Abstract

The use of learner modelling approaches is critical for providing adaptive support in educational computer games, with predictive learner modelling being among the key approaches. While adaptive supports have been shown to improve the effectiveness of educational games, improperly customized support can have negative effects on learning outcomes. To tackle these challenges, we present a novel approach, called DeepLM, that considers a series of time windows representing both sequences of learners’ actions during gameplay and estimation of their current competencies (using stealth assessment) to model learners and accordingly predict their future performance. The approach employs a variant of deep neural networks to early predict learners’ midterm and final scores simultaneously. The results show that using 20-50% of learners’ action sequences can early predict their final scores, with a cross-validated convolutional neural network (CNN) achieving an area under the curve (AUC) and accuracy of 0.879 and 85.3%, respectively. The same model can also achieve high accuracy in predicting midterm and final scores at the same time, with an AUC and accuracy of 0.848 and 77.9%. Overall, the CNN model outperforms recurrent neural network, long short-term memory, and baseline multilayer perceptron (MLP) models in predicting learners’ final performance and performs better than the baseline MLP model in predicting learners’ midterm and final performance using both cross-validation and independent datasets. These findings show the potential of the proposed approach in accurately early predicting learners’ performance, allowing educators and game designers to tailor interventions and support mechanisms that could lead to optimized learning outcomes.

https://doi.org/10.34105/j.kmel.2024.16.019