Cross-Platform Student Performance Prediction: A Multi-Task Learning Approach with Transferable Knowledge

Educational data mining faces significant challenges due to platform heterogeneity, where models trained on one platform often fail to generalize to others. We propose PAMNet (Platform-Aware Multi-task Network), a novel deep learning architecture that leverages shared representations while preserving platform-specific characteristics. Validated on the NeurIPS 2020 Education Challenge dataset ($\mathbf{1 5. 8 7 M}$ records) and ASSISTments 2009-2010 dataset ($\mathbf{1. 0 1 M}$ records), PAMNet achieves transfer efficiency rates of $\mathbf{7 9. 4 \%}$ and $\mathbf{6 6. 1 \%}$ in bidirectional cross-platform scenarios, significantly outperforming standard multi-task learning baselines (44.5% and 58.3%). Our error analysis reveals that PAMNet reduces critical false negative rates by up to 56.3%, demonstrating more educationally appropriate error distributions. Feature importance analysis shows high consistency (correlation $=0.71$) for proficiency-related features across platforms. These findings establish that explicit platform-aware modeling enables robust cross-platform deployment while maintaining high predictive accuracy, advancing toward truly transferable educational AI systems.