DeepFM-AE: Learning Student Representations for Performance Score Prediction

Abstract

Student performance prediction is essential for personalized learning support in online education. This paper proposes DeepFM-AE, a hybrid autoencoder architecture that learns comprehensive student representations for accurate performance score prediction. The model combines factorization machines (FM) to capture educationally meaningful feature interactions with deep neural networks (DNN) to learn complex nonlinear patterns, all within an autoencoder framework that ensures robust representation learning. Evaluated on the Open University Learning Analytics Dataset (OULAD) with 26,727 students, DeepFM-AE achieves a Mean Absolute Error (MAE) of 0.1554, representing a 4.78% improvement over the competitive DeepFM baseline (MAE=0.1632). While this entails a trade-off in other regression metrics, we establish MAE as the primary metric of interest for this educational task. Ablation studies reveal that the reconstruction constraint is crucial for this performance gain. The learned representations show clear semantic structure when visualized, validating the effectiveness of our dual-encoder design for educational data.