Student placement and skill ranking predictors for programming classes using class attitude, psychological scales, and code metrics

Abstract

In some situations, it is necessary to measure personal programming skills. For example, often students must be divided according to skill level and motivation to learn or companies recruiting employees must rank candidates by evaluating programming skills through programming tests, programming contests, etc. This process is burdensome because teachers and recruiters must prepare, implement, and evaluate a placement examination. This paper tries to predict the placement and ranking results of programming contests via machine learning without such an examination. Explanatory variables used for machine learning are classified into three categories: Psychological Scales, Programming Tasks, and Student-answered Questionnaires. The participants are university students enrolled in a Java programming class. One target variable is the placement result based on an examination by a teacher of a class and the ranking results of the programming contest. Our best classification model with a decision tree has an F-measure of 0.912, while our best ranking model with an SVM-rank has an nDCG of 0.962. In both prediction models, the best explanatory variable is from the Programming Task followed in order by Psychological Sale and Student-answered Questionnaire. Our classification model uses 9 explanatory variables, while our ranking model uses 20 explanatory variables. These include all three types of explanatory variables. The source code complexity, which is a source code metrics from Programming Task, shows best performance when the prediction uses only one explanatory variable. Contribution (1), this method can automate some of the teacher’s workload, which may improve educational quality and increase the number of acceptable students in the course. Contribution (2), this paper shows the potential of using difficult-to-formulate information for an evaluation such as a Psychological Scale is demonstrated. These are the contributions and implications of this paper.