Research On Attention-Based Dynamic Cognitive Diagnosis Method For Intelligent Education

Due to the widespread popularity of the Internet and the impact of the epidemic,online education has become one of the main ways for students to study.Online education platform is different from the traditional classroom teaching mode,students can choose course content for their own learning.In order to improve students’ learning efficiency and select suitable courses for teaching among many courses,the education platform needs to provide students with personalized teaching contents and learning routes according to their knowledge proficiency.Since students’ knowledge proficiency is constantly changing with the learning process,it is important to obtain students’ knowledge proficiency in real time.Cognitive diagnosis is an important core component of intelligent education systems,as it captures students’ proficiency of specific knowledge points based on their historical learning records and predicts their future learning performance.Current cognitive diagnosis approaches mainly use hand-designed functions（e.g.,logistic function）to simulate the linear interaction process of students in the learning process,however,this is not able to capture the complex relationships that students exhibit in the process of problem solving.With the development of deep neural networks,deep learning neural network-based knowledge tracking models as well as neural cognitive diagnosis models have been proposed.These models use multidimensional vectors as student factors and use deep neural networks to model the interaction function between student factors and exercise factors.Although each of these approaches has advantages,they all have some drawbacks at the same time.The traditional cognitive diagnosis methods,which have good interpretability,are too simple in their interaction functions and cannot be diagnosed dynamically,which do not match the dynamic change of students’ knowledge proficiency with the process of practice.In contrast,knowledge tracking methods can perform dynamic diagnosis compared with cognitive diagnostic methods and generally have better results,but at the same time lack better interpretability as in cognitive diagnosis methods.At the same time,these methods focus on the current exercise and ignore the connection with the past exercises.To address above problems,we propose an Attention-based dynamic Cognitive Diagnosis（ACD）.The main work is as follows.We divide the cognitive diagnosis process into two layers,student-exercise factor embedding layer and student-exercise factor interaction layer,based on the three main components of cognitive diagnosis,namely student factor,exercise factor,and interaction function between student factor and exercise factor.In the student-exercise factor embedding layer,the correlation weights between the exercises and knowledge concepts and the features of the exercises are extracted based on the exerciseknowledge concept association graph using graph attention network,and exercise factors are constructed by them.At the same time,we use the memory matrix as the student knowledge proficiency,and construct the student factors according to the correlation weights to achieve dynamic update of the student knowledge proficiency and finally constitute the student-exercise factor feature.In the layer of student-exercise factor interaction,we introduce a multi-head attention mechanism instead of a simple fully connected layer to extract the historical features of exercises and historical features of interactions in students’ historical records,and use these historical features,combined with forgetting characteristics,to simulate the process of students’ answering current exercise with reference to their past solving experience.Finally,in the prediction layer,we make predictions about students’ performances.To verify the validity of the model,we conducted ablation experiments and comparison experiments on three datasets,Junyi,Statics 2011,and ASSISTment 2012,and evaluated the prediction accuracy of the model using the predicted ACC and AUC.The results of the ablation experiments show that both the graph attention network and the self-attention network used in ACD can effectively improve the performance of the model;the results of the comparison experiments show that the ACD model proposed in this paper has better prediction accuracy compared with other comparison models.