Improved randomized learning algorithms for imbalanced and noisy educational data classification

Ming Li, Changqin Huang, Dianhui Wang, Qintai Hu, Jia Zhu, Yong Tang

Abstract: Despite that neural networks have demonstrated their good potential to be used in constructing learners which exhibit strong predictive performance, there are still some uncertainty issues that can greatly affect the effectiveness of the employed supervised learning algorithms, such as class imbalance and labeling errors (or class noise). Technically, imbalanced data resource can cause more difficulties or limitations for learning algorithms to distinguish different classes, while data with labeling errors can lead to an unreasonable problem formulation due to incorrect hypotheses. Indeed, noise and class imbalance are pervasive problems in the domain of educational data analytics. This study aims at developing improved randomized learning algorithms by investigating a novel type of cost function that focuses on the combined effects of class imbalance and class noise. Instead of concerning these uncertainty issues isolation, we present a convex combination of robust and imbalanced modelling objectives, contributing to a generalized formulation of weighted least squares problems by which the improved randomized learner models can be built. Our experimental study on several educational data classification tasks have verified the advantages of our proposed algorithms, in comparison with some existing methods that either takes no account of class imbalance and labeling errors, or merely consider one specific aspect in problem-solving.