Harnessing Bidirectional Learning for Enhanced Student Performance Prediction Model

Online learning behaviors significantly impact academic performance. Studies have shown that predicting student performance through learning behaviors can identify students' propensity to fail, helping teachers intervene in time to prevent dropouts. The rapid development of online education generates data characterized by high latitude, dynamics, and nonlinearity, posing a challenge to performance prediction modeling. However, most existing methods utilize unidirectional neural network structures, disregarding the joint influence of forward and backward associations of learning on student performance. Additionally, methods for predicting student performance often use characteristics such as demographics, learning behaviors, and assessments of learning. However, not all characteristic behaviors are closely related to academic performance, which emphasizes the importance of carefully selecting characteristics. This study presents the proposed Bidirectional Performance Prediction Model (BPPM), which includes the Feature Distiller and Random-Mask-BERT model. The Feature Distiller and Random-Mask mechanisms are used to enhance the model's bidirectional learning ability. Experimental verification shows that the BPPM can successfully predict student performance and outperform other baseline models. Finally, this study used the mRMR feature selection algorithm to eliminate redundant features and designed ablation experiments. It has been demonstrated that the use of Feature Selection, Feature Distiller, and Random-Mask mechanisms plays an irreplaceable role in improving the predictive performance of the model.