Predicting student academic performance from smartphone usage patterns requires careful preprocessing of heterogeneous mobile sensor data before deep learning model training. This research introduces AMIN (Adaptive Multimodal Imputation and Normalization), a systematic preprocessing framework designed to standardize noisy, incomplete smartphone behavioral data for educational prediction tasks. Through empirical evaluation across multiple deep learning architectures, we demonstrate that strategic preprocessing choices substantially impact model performance—often exceeding the improvements gained from architectural modifications alone. The AMIN framework integrates temporal-aware imputation for missing values, modality-specific normalization tailored to different sensor types, and per-student baseline adjustment to prevent identity-based shortcuts in learning. Comparative analysis shows AMIN achieving performance improvements of 8-14% over conventional preprocessing approaches across MLP, LSTM, and Bi-LSTM architectures. This work establishes a reproducible baseline preprocessing methodology that enhances model stability, enables fair architectural comparisons, and facilitates adoption of standardized practices in educational data science research.
