Deep Knowledge Tracking Based On Students' Answer Sequence

Knowledge tracking is one of the important research topics in educational data mining.The knowledge tracking task is to track the student's knowledge state according to the students' historical learning behavior,then predict how the student will behave in future interactions.By modeling the student's knowledge state,it can provide students with personalized learning guidance,avoid students to spend much time doing endless exercises,and also help teachers to better understand the students' learning level and adjust the teaching plan accordingly.This thesis takes the students' answer sequence data as the research object,and mainly carries out the following related research work for deep knowledge tracking:(1)Aiming at the problem that the existing deep knowledge tracking methods do not fully consider the impact of students' behavior characteristics on the students' learning process,a deep knowledge tracking model based on feature extraction(DKT-LDA)is proposed.The model uses the linear discriminant analysis method to automatically extract the potential information from the students' multiple behavior characteristics and learn their representations,fully integrating the impact of each feature on student learning performance and reducing the feature dimension.The model performed the knowledge tracking task on the ASSISTments09 dataset and obtained an AUC value of 80.92%,which is better than the current mainstream knowledge tracking algorithms.(2)Aiming at the problem that the existing deep knowledge tracking methods do not take into account the impact of student abilities and the exercises difficulty on the learning process of students,a deep knowledge tracking model based on student ability and exercises difficulty fusion attention machine(DARNNA)is proposed.This model uses the students' answer records to quantify students' ability and exercises difficulty,and then obtains its embedding representation through a neural network,and then combines with answer records to reconstruct the model input,and finally incorporates attention machine to make it pay more attention to interactive records with similar student abilities and exercises difficulty when predicting.This model obtained AUC values of 83.71% and 78.55% on the ASSISTments09 and ASSISTments15 datasets.The results are better than the existing publicly published knowledge tracking models.(3)Aiming at the existing deep knowledge tracking models do not take into account the slip and guess in the process of doing the questions,a deep knowledge tracking model that combines slipping rate and guessing rate(DKT-GS)is proposed.The model first defines and calculates the slipping rate and the guessing rate,and then integrates the slipping rate and the guessing rate to improve the hidden layer of the LSTM network,and models the real learning state of the students,so that it can better predict the students' behavior on the next question.This model obtained AUC values of 85.11% and 79.17% on the ASSISTments09 and ASSISTments15 datasets.The results are better than the existing advanced knowledge tracking models.In summary,the model proposed in this thesis can more accurately model the student's knowledge state,and at the same time improve the accuracy of predicting student performance,promote the personalized development of students,and effectively avoid information tragedy and overload.