Research On Deep Knowledge Tracing Model Based On Feature Fusion And Network Embedding

The popularity of online education has produced a huge amount of educational data,making it difficult for teachers to know how each student is learning and provide personalized guidance.As an important subject of education data mining,knowledge tracing can model students’ knowledge state according to their historical answer records,so as to provide personalized learning experience for each student.Among all knowledge tracing methods,the deep knowledge tracing(DKT)model has achieved good results due to the high flexibility of neural networks and is currently the most widely used method.However,the DKT model only predicts student performance based on their answer sequences,ignoring the influence of student behavior characteristics on their accuracy.Moreover,due to the special nature of online learning,student answer data is often large and sparse,making it difficult to achieve good predictive performance on datasets with insufficient effective data.To address these problems,this thesis fully considers and explores various behavior characteristics to improve model performance,captures the correlation between exercises to alleviate the impact of sparse data,and proposes two new deep knowledge tracing models based on feature fusion and network embedding methods.The main research work is as follows:(1)Propose a deep knowledge tracing model based on the fusion of student behavior features.First,various behavior features generated by students in the classroom learning,forum discussions,and answering interaction processes are excavated.The cross-hot encoding method is used to establish the relationship between various features and exercises,and the fused feature vector is obtained.Then,the features are input into a pre-trained autoencoder to obtain the input feature sequence after dimensionality reduction.Finally,the long short-term memory network is used to replace the recurrent neural network used in traditional DKT to alleviate the problems of gradient explosion and vanishing.To prove the effectiveness of the model,a large number of experiments were conducted on the MOOCCube X dataset,and the experimental results showed that the model outperformed multiple comparison methods and could better predict student future performance.(2)Propose a deep knowledge tracing model based on question network embedding.First,excavate the relationship among students,knowledge points and questions to construct question network.Node2 vec is used to capture the potential correlation between questions from the global structure of the network,and the vector representation of questions is obtained.Then,the attention mechanism is used to capture the relationship between the historical answer sequence and the current exercise,and the exercise most relevant to the current state in the historical sequence is reinforced.Finally,the exercise vector is fused with the feature vector of the hidden layer state at time t to output the final prediction result.A large number of comparative experiments on the MOOCCube X 、 ASSISTments2009 、ASSISTments2012 、 ASSISTments2015 、 Ed Net datasets have verified that the model can better alleviate the impact of sparsity on the model and improve the performance of the original model.