Programming Knowledge Tracing Based On Deep Learning

With the development of the Internet,online education is becoming more and more popular,and the massive learning data generated by online education platforms provide a database for education researchers to explore how to better assist students in learning.Implicit information,such as students’ learning behavior pattern and knowledge level,is contained in this massive data.Knowledge Tracing(KT)is a research hotspot in recent years.It assesses students’ knowledge level and predicts their future performance according to their historical learning records,usually by constructing student models.In 2015,deep neural network was applied in Knowledge Tracing for the first t ime.A number of deep Knowledge Tracing models have been proposed and applied in many practical teaching scenarios since then.Now we are in the era of “Internet of Everything”,programming skills have gradually become one of the necessary skills for people,and programming education tends to start at a younger age.Therefore,research on programming education begins to appear.Programming Knowledge Tracing(PKT)is an application of deep Knowledge Tracing in programming learning.Programming exercises differ from other exercises in that:1)There is usually no single standard answer to an exercise,and multiple knowledge concepts are inspected at the same time in one exercise.Students can write programs to answer open-ended questions based on what they have learned.2)The source code of students’ answers contains their grasp of the knowledge concepts inspected by the exercises.3)Students may need to attempt on the same exercise many times,modify and submit the code constantly to pass the exercise.In view of the behavioral particularity of programming learning,this paper proposes two models,with the main contributions as follows:1.Considering the relationship between the source code submitted by students and the knowledge concepts,and the multi-concept characteristic of the exercise,a programming knowledge tracing model PKT-ATTN with a code-concept interaction module is proposed.The main structure is the LSTM network,added with a code-concept interaction module implemented by attention mechanism,which is used to search the code patterns corresponding to the features of knowledge concepts,and explore how students use the knowledge concepts comprehensively in the code,in order to better update students’ knowledge level and predict their subsequent problem solving performance.2.Considering the iterative code feature of students’ repeated attempts on the same exercise,a programming knowledge tracing model PKT-CNN integrating 3D convolution module was proposed.The main structure is also LSTM.We first represent the abstract syntax tree of the code with a 2D matrix,then stack the code representation matrix of the student’s several most recent attempts on the same exercise,and get the 3D tensor representing the changing process of these submitted codes.A 3D convolution module is used to mine the iterative code features,to assist the prediction of students’ future performance.Both models proposed in this paper are based on the classic deep Knowledge Tracing model DKT,with different code feature fusion modules a dded.Experiments show that the two modules can improve the performance of programming knowledge tracing models.Although DKT is used as the basic model,the proposed code feature fusion modules can be easily transferred to other deep Knowledge Tracing models to make them suitable for the specific teaching scenario of programming education.