线上线下混合式教学中的学业预警研究

Research on Academic Early Warning in Blended Online and Offline Teaching

摘要: 随着在线教学平台的普及和学习分析技术的发展，教育大数据的应用在评估线上课程学习效果中日益凸显其重要性。本研究旨在识别影响线上课程学生学习效果的关键特征，建立学习效果预测模型，实现对线上课程学生学习效果的综合评价。在特征选择上，本文利用相关系数变量选择、互信息变量选择、基于随机森林递归式变量选择、Lasso嵌入法和SHAP法这五种方法对特征进行选择分析，最终利用TOPSIS综合筛选特征，最终筛选出8个变量参与模型的预测，综合五种方法使得我们能够全面、准确地识别出影响学习效果的关键特征。在预测模型构建上，本文选取随机森林、LightGBM以及BP神经网络三种经典机器学习与深度学习模型，对学生学习成绩进行量化预测，并采用粒子群算法对各模型进行超参数优化，提升模型拟合与泛化能力。为客观量化模型性能，本研究选取均方误差与拟合优度作为核心评价指标，通过对比分析三种模型的预测效果，结果显示BP神经网络算法在均方误差、拟合优度及整体预测稳定性方面均表现最优。研究结果表明，神经网络模型成功地评估了线上课程学生的学习效果，其预测准确性和解释能力为在线教育的发展和个性化教学提供了重要的理论依据和实践指导，可以帮助教育者识别出哪些教学资源和学习活动对学生的学习效果有显著影响，从而能够进行更加有针对性的教学干预，优化教学策略，最终提高学生的学习成绩。

Abstract: With the widespread adoption of online teaching platforms and the advancement of learning analytics technology, the application of educational big data has increasingly highlighted its importance in assessing the learning outcomes of online courses. This study aims to identify key features influencing student performance in online courses, establish a predictive model for learning outcomes, and achieve a comprehensive evaluation of student learning effectiveness in online courses. In terms of feature selection, this paper employs five methods—correlation coefficient-based variable selection, mutual information-based variable selection, recursive feature elimination based on random forests, Lasso embedding, and SHAP—to analyze and select features. The TOPSIS method is then used for comprehensive feature screening, ultimately identifying eight variables for model prediction. By integrating these five approaches, we achieve a thorough and accurate identification of key features influencing learning outcomes. For predictive model construction, this study selects three classical machine learning and deep learning models—Random Forest, LightGBM, and BP neural networks—to quantify student academic performance. Particle swarm optimization is applied to fine-tune hyperparameters for each model, enhancing both fitting and generalization capabilities. To objectively evaluate model performance, mean squared error and coefficient of determination are chosen as core metrics. Comparative analysis of the three models’ predictive effectiveness reveals that the BP neural network algorithm demonstrates superior performance in terms of mean squared error, coefficient of determination, and overall predictive stability. The research findings indicate that the neural network model successfully evaluated the learning outcomes of students in online courses. Its predictive accuracy and interpretative capabilities provide crucial theoretical foundations and practical guidance for the development of online education and personalized teaching. This model helps educators identify which teaching resources and learning activities significantly impact student performance, enabling more targeted instructional interventions, optimizing teaching strategies, and ultimately improving student academic achievement.