Key attribute for predicting student academic performance

Abstract

Predicting student final score from student's attributes is an important issue of learning analytic. Not only to achieve high prediction performance but also to identifying the key attributes is an important research theme. This paper evaluated exhaustively the prediction performance based on all possible combinations of four types of attributes - behavioral features, demographic features, academic background, and parent participation. The behavioral features are given as numerical data. But, we represented them as pair of an attribute name and the value. This vectorization yields 417 dimensional data, while naively represented data has 68 dimension. By applyig support vector machine and feature selection, we obtained the optimal prediction performance, with respect to feature selection, with accuracy 0.8096 and F-measure 0.7726. We confirmed that the behavioral feature is so crucial that the accuracy reaches 0.7905 without other features except behavioral feature. The combination of behavior feature and demographic feature gained F-measure 0.7662.