Construction And Application Of Knowledge Tracking Model Based On Problem Composition

Knowledge tracking is based on students’ behavior sequence to model and predict students’ understanding of knowledge.It is the core technology when building intelligent education systems.In smart education systems,the first step is to predict students’understanding of knowledge,before performing accurate push,planning students’ learning path or building a library of test points.Traditional knowledge tracking models include Bayesian knowledge tracking and knowledge tracking models based on Probabilistic Matrix Factorization.BKT has poor long-term dependency and PFM has poor interpretability.In recent years,scholars have proposed a knowledge tracking model based on the theory of deep learning.Compared with traditional knowledge tracking models,knowledge tracking models based on deep learning improves accuracy of prediction results.This paper also constructs a knowledge tracking model based on deep learning and proposes a novel knowledge tracking model SWKT based on problem composition.The major work of this dissertation can be summarized as follows:(1)Existing models do not take into account the background information of a specific problem to extract features of data,such as the potential difficulty level of the question,the accuracy rate of students’ answers,and the speed of students’ responses.The accuracy rate of prediction is low since the result is simply based on students’ understanding of test points to predict whether students will answer questions correctly in future.When interactions between students and questions are not sufficient,the prediction result is not ideal.Therefore,background information of questions is very necessary for enhancing identification accuracy of KT model at a more fine-grained level.As sparse features are taken as input by existing models,this dissertation proposes a edge information(PBEI)embedding method based on pre-training by correlating three factors of questions,test points of questions and students’ information.PBEI considers correlation among different test points of questions,correlation among combination of test points of questions,correlation among different questions,the characteristic degree of each question,the characteristic degree of each test point,each student’s understanding of all test points to be used as the feature vector to improve sparsity of features extracted by existing technologies.(2)In practice,students answer a large number of similar questions with varied difficulty levels.The existing models do not take this factor into account when making predictions,which will also affect their accuracy rate for predicting students’ answers to the next test point.By structures of questions,this dissertation divides one question into two parts:data and the corresponding test point,and computes the weights of the features of questions and test points in a way similar to the attention mechanism.The weighted summation method is used to combine content of one question and the characteristic of the corresponding test point’s importance level for construction of the knowledge tracking model.Hence the prediction efficiency of the proposed model is improved when predicting students answers to the next test point.This dissertation evaluates the accuracy rate through two metrics:ACC and AUC.Compared with existing models,experiments show that SWKT achieves improvement.(3)Due to the use of sparse features,existing models often rely on complex neural network structures to predict and yield reliable results with relatively low efficiency.In the SWKT model,a lot of feature vectors are introduced in the early stage,and the efficient linear model is used by the model for final prediction.At the same time,the linear model adopts only two kinds of parameters for training,namely the weight parameters and bias parameters.By this way,the efficiency of the proposed model has been improved,and experiments show that the system efficiency has been greatly reduced.