AI-Driven Personalized Learning Effect Prediction Based on Machine Learning Algorithms

Abstract

Generative and learning analytics artificial intelligence (AI) is rapidly entering middle school classrooms, expected to enhance personalized learning, improve formative assessment, and reduce the burden on teachers. Australia has formed a relatively clear ecosystem at the policy and curriculum levels: The federal education department has released the "Generative AI Framework for Australian Schools", putting forward principles such as safety, ethics, transparency and capacity building; In the V9 version of the ACARA course, resources for the connection between AI and disciplinary capabilities are provided. NAPLAN has fully shifted to online adaptive evaluation, providing the institutional and technical foundation for personalized assessment. Meanwhile, UNESCO advocates "people-oriented" generative AI education governance, while the OECD emphasizes the value of learning analytics and data-driven adoption. To address the limitations of existing algorithms, this paper proposes a regression prediction algorithm based on multi-head attention mechanism to optimize the bidirectional long short-term memory network (BiLSTM). The study first conducted a correlation analysis, and at the same time selected Decision tree, Random Forest, AdaBoost, CatBoost and LightGBM as comparison models to carry out experiments. From the perspective of performance indicators, the proposed model performs the best in all dimensions and comprehensively outperforms the comparison models. Among the error-related indicators, its MSE (31.688) is approximately 21.7% and 22.0% lower than LightGBM (40.491) and AdaBoost (40.604), respectively, providing better control over large errors. RMSE (5.629) and MAE (4.616) were also the smallest, decreasing by 11.5%-11.7% and 8.1%-11.6% respectively compared with the two models, and the overall deviation was smaller. The MAPE (6.436) is 8.0%-12.5% lower than that of the two models, with small relative errors and stable accuracy. In terms of goodness of fit, its R (0.765) is 6.7%-11.8% higher than that of the two models, which can explain 76.5% of the variation of the dependent variable and has a stronger ability to capture data patterns. This model provides reliable technical support for AI education platforms to accurately analyze students' learning status and optimize personalized teaching strategies, which has significant practical significance for promoting the efficient implementation of personalized education.