Multimodal Student Performance Prediction Based on Integrated Learning Feedback

Performance prediction for students is of paramount importance in school teaching and management, providing valuable guidance for personalized instruction and pedagogical interventions. We employ word2vec (a technique for converting text into a vector representation) for text embedding and utilize supervised models to predict final performance. The study compares feature-level fusion, which combines different data types before model training, with decision-level fusion, which makes predictions from individual models and then combines them. We also delve into early performance prediction. Experimental results can be demonstrated. (1) In terms of prediction effectiveness for performance prediction, unit tests lead the way with an RMSE of 0.2155, followed closely by experiments at 0.2206. Homework assignments rank next with an RMSE of 0.2402, and self-feedback bring up the rear with an RMSE of 0.2454. (2) The two fusion methods exhibit comparable predictive power: feature-level fusion has an RMSE of 0.2462, while decision-level fusion achieves a slightly lower RMSE of 0.2313. (3) The integration of unit tests with self-feedback significantly enhances the precision of performance prediction. For decision-level fusion, the RMSE shows a steady decline from 0.25 to 0.245, and for feature-level fusion, the RMSE also demonstrates a consistent improvement, dropping from 0.244 to 0.234.