MultiFAR: Multidimensional information fusion with attention-driven representation learning for student performance...

Area of Science:

• Educational Technology
• Machine Learning
• Data Science

Background:

• Online learning platforms and advanced computing enable remote education.
• Existing student performance prediction models use basic deep learning with limited features.
• There's a need for sophisticated models to analyze multifaceted student data.

Purpose of the Study:

• To propose MultIFAR, an attention-driven deep learning model for efficient student representation learning.
• To integrate multi-dimensional student data including demographics, assessments, and VLE interactions.
• To improve the accuracy and early prediction of student performance and risk.

Main Methods:

• Developed MultIFAR, integrating bidirectional LSTM and CNN with an attention mechanism.
• Utilized demographic, assessment, and VLE interaction data for comprehensive student behavior analysis.
• Employed the Open University Learning Analytics (OULA) dataset for empirical evaluation.

Main Results:

• MultIFAR achieved high accuracy (80.31%–97.12%) in student performance prediction, outperforming baseline methods.
• Ablation analysis showed diurnal interaction data has the most significant impact, while demographics have the least.
• The model successfully extended to early prediction of at-risk and high-performing students.

Conclusions:

• MultIFAR demonstrates superior efficacy in student performance prediction and early risk identification.
• Integrating multi-dimensional data and advanced deep learning techniques is crucial for educational analytics.
• The model's flexibility allows for adaptation to various educational prediction tasks.