The Research And Application Of Modeling Students' Ability In Personalized Education System

With the further development of educational informationization in China,many personalized learning systems have been launched.Personalized education systems collect massive learning process data of students,on one hand,to provide students with informative diagnostic reports and targeted learning resources;on the other hand,to support their teaching decision making and enhance teaching efficiency.The most fundamental and challenging task of personalized education systems is modeling students' ability.Inspired by the successful applications of artificial intelligence and big data technologies in fields such as medical,finance etc.,researchers have introduced those interdisciplinary methods to the task of modeling students' ability.Specific challenges of the task are as follows:1)limited information for modeling could be used due to the data gathering difficulties and data sparsity.2)Poor interpretability of traditional machine learning models plays a limited role on this task.3)Integration with interdisciplinary knowledge and domain-specific knowledge is difficult.Aimed at those difficulties mentioned above,in this thesis,we will conduct a deep study in modeling students' ability task,and the contributions of this thesis are concluded as the follows.First,this thesis proposed a novel Exercise-Enhanced Recurrent Neural Network(EERNN)framework for modeling students' ability.Accuracy of students' performance prediction on specific test items(or exercises)is an important measure of' the students'ability models.Existing methods mostly just exploit the historical answer records of students,or regard each exercise as knowledge concepts representation,causing server information loss.In this thesis,we proposed a novel Exercise-Enhanced Recurrent Neural Network(EERNN)to take full advantage of both students' answer records and the exercises' texts.Specifically,to learn each exercise representation,we first designed a bidirectional LSTM from its text description without any expertise and information loss.Then,to track students' ability states,we proposed a new LSTM based architecture in students' sequential exercising process with the combination of exercise representations and their performances on these exercises.Finally,to predict students'performance on new test items,we proposed two strategies under EERNN,i.e.EERNNM with Markov property and EERNNA with Attention mechanism.It is worth mentioning that EERNNA assumes that students' representations should be biased in the view of different exercise representations,which helps the model to overcome long-term answer record sequences and track the most focused ability states for predictions on various kinds of exercises.EERNNA outperforms other related models by a big margin if we have a large amount of training data,i.e.exercises and answer records.To a certain extent,EERNN can also deal with data sparsity and cold start problem.In the scenes of training data shortage(specially,when a new system is just online),we need to explore new methods,which will be discussed in the next section.Second,this thesis proposed a knowledge graph based students' ability modeling method.To solve the data sparsity problem and increase model interpretability for making the recommendations more reasonable,this thesis proposed to integrate educational knowledge graph into deep learning models for modeling students' ability.We also regarded the accuracy of students' performance prediction as the measure of the students' ability models.In this framework,we used the knowledge graphs as a carrier of education domain-specific knowledges,i.e.typical exercises(anchor exercises)are denoted by the graph nodes and the difficulty partial orders of exercises are denoted by graph edges.Specifically,in the modeling process,we proposed to use auto-encoders with the combination of knowledges in the anchor graph.By taking advantages of the graph knowledge,this model can ease the data sparsity problem and increase model interpretability.Moreover,auto-encoders could mine the hidden relations between students and exercises,exercises and exercises,thus improve the prediction accuracy.This framework has significant advantages in both modeling and applications in scene of knowledges graph based education recommendation systems.Third,this thesis proposed a typical student prototypes mining method.Lamination teaching method is a new rising personalized education method,in which teachers group the students according to the students' characteristics and teach accordingly.To assist lamination teaching,mining typical student prototypes automatically from students learning process data becomes an urgent demand.Moreover,a real-world student should be represented by several prototypes.This thesis proposed a novel convex dictionary learning model to address this issue,which restricts that the learned dictionary elements must be in a convex hull of the original data objects.This model makes the virtual learned prototypes more close to some real examples,on one hand,ensuring the accuracy of data reconstruction and enhancing category information of original data;on the other hand,improving the interpretability of the mined typical prototypes.We also developed an efficient training algorithm for this model and provided a strict proof of its convergence.Experiments on both educational datasets and image datasets showed promised quality of the mined typical prototypes,which indicated that convex dictionary learning is a general method in various fields with good research potential.Fourth,the methods proposed in this thesis can be well applied in modern online education systems.Besides theoretical research and experimental analysis,in each chapter,this thesis also described how to apply these researches into practical applications in detail,such as Test Paper Generating System,Personalized Learning System,Lamination Teaching System.These successful applications verified the effectiveness and their practical value of the proposed models.We hope that the elaboration of specific applications could attract the attention of both researchers and engineers to promote the development of this direction.