Iterative and Statistical Analytical Review of Predictive Modeling Approaches in Educational Systems: A Comprehensive Benchmark of AI-Driven Methods

Abstract

The rapid digitization of education and the growing availability of multimodal student data have driven the development of advanced AI-based educational prediction systems. The proliferation of algorithms, feature spaces, and performance metrics has resulted in fragmented insights and a lack of consensus on optimal modeling practices. This study presents a systematic analytical review of landmark research on predictive modeling in education, covering student performance, dropout risk, employability, mental health, and engagement. A robust benchmarking framework is established to evaluate models across six key performance metrics: root mean square error (RMSE), accuracy, latency, computational complexity, precision, and recall. The review examines hybrid deep learning architectures (e.g., CSSA-Deep RNN, KANFormer, CNN-GRU), metaheuristic-based approaches (PSO, ACO, t-SIDSBO), interpretable models (SHAP, LIME, federated learning), and affective computing systems incorporating facial emotion recognition (FER) and natural language processing (NLP). To enable consistent comparisons, numerical performance scores were standardized and systematically reported. The findings indicate that while accuracy remains a primary concern, practical deployment necessitates careful trade-offs among complexity, latency, and interpretability. This study addresses critical gaps in prior reviews by emphasizing model explainability, robustness, and context-awareness. Future research directions are identified, including multimodal fusion, edge deployment, longitudinal modeling, and causal explainability. This work serves as a resource for researchers, policymakers, and educational technology developers in designing, evaluating, and deploying intelligent educational systems.