Robust personalized representation learning for sparse user modeling

Modeling individualized behavior in sparse and noisy environments remains a core challenge for personalized systems, especially in educational contexts like Knowledge Tracing (KT). We propose PRL-SUM (Robust Personalized Representation Learning for Sparse User Modeling), a novel framework that leverages causal graph modeling and adversarial learning to extract high-dimensional user representations from interaction histories indexed by user IDs, eliminating the need for demographic data or manual annotations. PRL-SUM comprises two key modules: (1) a User Autoencoder that builds causal behavior graphs and learns robust embeddings via a generator-discriminator mechanism; (2) a dynamic performance predictor combining attention-based knowledge state modeling with a gated multi-expert network for adaptive performance inference. Unlike traditional KT models assuming user homogeneity, PRL-SUM effectively disentangles individual signals from noise, improving generalization under data sparsity. Extensive experiments on six datasets (five public KT benchmarks and one proprietary dataset) show PRL-SUM consistently outperforms state-of-the-art baselines, achieving up to 8.5% AUC gain and 23.8% root mean square error (RMSE) reduction on CompArch. On average, it improves AUC by 3.6%, ACC by 2.2%, and reduces RMSE by 9.0%, while maintaining robustness under varying noise and data retention levels. These results underscore PRL-SUM’s effectiveness in robust, scalable personalized modeling, with broad applicability in education, recommendation, and behavioral analytics.