A Neural Network-Based Prediction Model for Students' Mathematical Metacognitive Monitoring Ability

Abstract

Under the strategic goal of advancing education-driven national development, enhancing students' mathematical metacognitive monitoring ability—defined as their capacity to monitor, regulate, and control cognitive processes—has emerged as a critical pathway to optimize mathematics learning efficacy. However, existing prediction methods suffer from two limitations: overreliance on subjective assessments while neglecting static student attributes, and insufficient modeling capacity of traditional machine learning for complex interdependencies. To address these challenges, this study integrates authoritative questionnaire designs to collect 1,816 valid student datasets, constructing a mathematical metacognitive monitoring dataset encompassing 21 static attributes. We propose SMML-Net, a deep learning-based prediction model that innovatively encodes static attributes into digital profiles and leverages recurrent-convolutional neural networks to extract comprehensive relational features among attributes, thereby overcoming the inefficient utilization of static data in conventional approaches. Experimental results demonstrate that SMML-Net significantly outperforms traditional machine learning models in prediction accuracy and robustness. Further feature importance analysis and chi-square testing identify ten critical static variables influencing metacognitive monitoring (e.g., learning persistence, self-regulation frequency). Based on these findings, we propose a tiered instructional strategy: foundational skill reinforcement for low-monitoring students, reflective guidance for intermediate learners, and strategy transfer emphasis for high-monitoring cohorts.