The impact of CNN MHAM-enhanced WRF and BPNN models for user behavior prediction.

Abstract

To address the challenge of user behavior prediction on artificial intelligence (AI)-based online education platforms, this study proposes a novel ensemble model. The model combines the strengths of Convolutional Neural Network (CNN), Multi-Head Attention Mechanism (MHAM), Weighted Random Forest (WRF), and Back Propagation Neural Network (BPNN), forming an integrated architecture that enhances WRF and BPNN with CNN and MHAM. Experimental results demonstrate that the improved BPNN model, when combined with WRF, outperforms individual models in predicting user behavior. Specifically, the integrated model achieves a prediction accuracy of 92.3% on the test dataset-approximately 5% higher than that of the traditional BPNN. For imbalanced datasets, it attains a recall rate of 89.7%, significantly surpassing the unweighted random forest's 82.4%. The model also achieves an F1-score of 90.8%, reflecting strong overall performance in terms of both precision and recall. Overall, the proposed method effectively leverages the classification capabilities of WRF and the nonlinear fitting power of BPNN, substantially enhancing the accuracy and reliability of user behavior prediction, and offering valuable support for optimizing AI-driven online education platforms.