融入课程知识图谱的KMAKT预测OA北大核心CSTPCD

Most existing deep knowledge tracking models have weak interpretability of knowledge tracking results and overlook the impact of the inherent correlation between exercises and knowledge points on the effectiveness of knowledge tracking and prediction results.To address these issues,this study proposes a course Knowledge graph and Multi-head Attention mechanism-based Knowledge Tracing(KMAKT)model to predict student performance.First,Word2Vec and Bidirectional Long Short-Term Memory(BiLSTM)networks are used to convert exercise answering sequence data into dense low dimensional vectors.The graph embedding model TransR is used to embed the course knowledge graph representation,and a multi-head attention mechanism is used to calculate the contribution of past exercise answering sequences to the current knowledge state.Subsequently,the influence of the precursor knowledge on the prediction results is explored using attention networks.Finally,prediction results are obtained using multi-layer neural networks,and the interpretability and prediction accuracy of the model are improved.The experimental results show that on the ASSISTments2017 dataset,the Area Under receiver operating Characteristic(AUC),accuracy,and F1 value of the KMAKT model are improved by approximately 5.20,4.20,and 2.40 percentage points,respectively,as compared to conventional Deep Knowledge Tracking(DKT).Therefore,The KMAKT model has good prediction performance.The visualization results of knowledge tracking on the Hunan University Signal and System(HNU_SYS)sub-dataset verify that the KMAKT model conform to the cognitive laws of education and have a certain degree of interpretability.