Deep Knowledge Tracing Model Of Elementary Mathematics Knowledge Points

In recent years,with the vigorous development of Internet education,a series of various online learning platforms and teaching systems have followed.However,both the traditional offline teaching mode and the existing online teaching scheme will be restricted by limited teacher resources,resulting in the inability of professors to provide personalized teaching guidance and learners to obtain personalized learning services.The proposal of intelligent education system is to automatically identify whether the learner has mastered the knowledge points or not in the process of interaction with the learner like a teacher and determine what and how to teach next.At present,this process is mainly realized by establishing a learner model.Knowledge tracing is one of the mainstream methods used for learner modeling.It aims to determine the learning attributes implicated in the learning process of learners by observing the accuracy of their answers to exercises and the time taken to answer questions,etc.over time,so as to dynamically predict the mastery of knowledge points of learners.The main difficulties of the knowledge tracing model are to simulate the complexity of human brain in different scenes and the complexity of knowledge among different knowledge points.Most traditional knowledge tracing methods and specific models require education experts to annotate labels in advance and only focus on exercises with a single knowledge point.Thanks to the development of deep learning,the main research object of this article is a new method called Deep Knowledge Tracing(DKT)proposed in recent years.Its main motivation is that it can reflect long-term knowledge relations,does not need the advance annotation by education experts and models the connection between complex knowledge points.However,compared with the actual learning situation,the original DKT model still has some obvious problems:1.The complexity of the neural network leads to the output results that cannot be reasonably explained to a large extent;2.In the time range when knowledge points can be mastered continuously,the prediction results of DKT model on the same knowledge points show no trend and fluctuate repeatedly;3.Normally,DKT model cannot produce better prediction results for training sets with small data volume.This thesis will start from the knowledge tracing model based on the Recurrent Neural Net-work(RNN)and its variant structure Long Short-Term Memory network(LSTM),study how to improve the prediction performance of the model and try to solve the above-mentioned targeted problems.First of all,the original DKT model only inputs the characteristics of ex-ercise labels and result labels,which is not in line with the actual application scenarios and may affect the output results of the model.Therefore,it is considered to introduce more other important features to form feature crosses for modeling so as to improve the accuracy of the model,and to find a dimensionality reduction method for the change of high-dimensional input caused by multiple features,so as to thoroughly optimize the network structure of the model.Secondly,aiming at the problem that the output of the learner's grasp of the same knowledge points in the original model presents abnormal and repeated fluctuations,it is decided to add appropriate regularization terms on the premise of introducing multiple fea-tures.By optimizing the objective function of the model to change the training strategy of the model,the problem is alleviated to a certain extent.Finally,aiming at the problems of DKT model's poor performance in small data sets and poor interpretation of output results,Bayesian network and LSTM network are combined into the same model by introducing fusion mechanism to solve them.This three research methods have all improved to differ-ent degrees on the basis of the prediction performance of the original model,and have also adopted different means of optimization for the main problems on the basis of the original model.The second method will also be implemented based on the first method.Therefore,this thesis is of great significance to the research on the original model of Deep Knowledge Tracing.