A Research On Feature Extraction And Feature Selection Of Programming Process For Programming Education

Predicting course grades based on programming process data is one of the two main research directions in the area of the research on programming process for programming education.The purpose of the research is to assist educators in identifying students with risk at the early stage of the course.Educators can provide additional for these students.This will help to reduce course failing rate and improve programming teaching.The feature extraction and feature selection of the programming process are the core work of this research.The quality of the programming process features selected by the researcher determines the upper limit of the accuracy of the course performance prediction.In the current research,the researchers did not take full advantage of programming process information when constructing the programming process features related to course grades.The main work of this paper with regard to this issue is as follows:(1)This paper summarizes the programming process data used in the existing programming process researches,and defines 15 kinds of programming process data with analytical value based on the research results and heuristic ideas.This paper implemented a well-designed programming process data collection system.(2)Using a variety of observation angles,this paper extracts 28 features that may be related to the course performance from the programming process data based on heuristic ideas,and rules out three features that are not related to the course performance based on data observation and statistical analysis methods.The remaining 25 programmatic process features related to course performance are used to construct the course performance predict model.Compared to the NPSM method,the experimental results show that the course performance prediction model based on these 25 programming process features has better prediction results.(3)This paper uses four feature selection methods to filter out features with low correlation with course performance from extract 25 programming process features from(2).According to the factors that the SFFS feature selection method can only select features based on a single training dataset and it needs researches to determine the target number of features,this paper propses a dynamic SFFS feature selection method based on multi-training dataset.Based on the dynamic SFFS feature selection method based on multi-training dataset,12 programming process features was selected which are better than the features choosed by the other feature selection methods.The prediction model based on these 12 features has a significant improvement in the prediction accuracy compared to the prediction model based on all 25 features.Therefore,this paper presents 12 programming process features that are significantly associated with course performance.