Research On Long Memory Time Convolutional Knowledge Tracking Model Based On Retention Rate

At present,the main challenge of online education is how to achieve "personalization",that is,how to accurately analyze students’ learning trajectories,and then recommend the learning content students need from the massive learning data and customize personalized learning courses.The main task of knowledge tracking is to grasp students’ knowledge mastery and predict students’ future answering performance by analyzing the data generated by students’ learning process.Knowledge tracking models based on deep learning are the current research hotspots,but deep knowledge tracing methods(DKT)based on recurrent neural networks will produce gradient disappearance and memory loss problems,which need to be further solved.In response to this problem,the following work has been done:(1)A network model that integrates time convolutional network and attention mechanism is proposed,and then successfully applied to knowledge tracking tasks.The model is a Time Convolutional Knowledge Tracking with Long Memory(LMKT).LMKT’s time convolutional network can capture answer records with a longer historical time to alleviate the memory loss problem in DKT.The attention mechanism can make the network automatically assign different importance to different questions,thereby emphasizing important answer information and ignoring secondary answer information.The experimental results are compared with three classical knowledge tracking models,showing that LMKT has higher accuracy and faster running speed.This shows that it can more accurately assess students’ knowledge mastery status and more accurately predict students’ future performance in answering questions.(2)Considering the different knowledge retention of students in the learning process,the knowledge retention module was constructed and successfully integrated into LMKT,and a Long Memory Time Convolutional Knowledge Tracking Model Based on Retention Rate(LMKT+RR)was proposed.The Knowledge Retention Rate module takes into account four main factors that affect students’ knowledge retention in the learning process:the number of answers to questions,the difficulty of comprehensive questions,the degree of differentiation of topics and the original knowledge level of students.Update the change in students’ knowledge level after answering questions at the current moment to the next moment through the retention rate module.LMKT+RR achieves better prediction performance than LMKT on all three datasets.This shows that the model can not only successfully take advantage of the advantages of time convolutional networks in long sequence modeling tasks,but also track the changes of students’ knowledge level in the learning process,so that it has better prediction performance.