Balancing sequential data to predict students at-risk using adversarial networks

Abstract

Class imbalance is a challenging problem especially in a supervised learning setup, as most classification algorithms are designed for balanced class distributions. Although various up-sampling approaches exist for eliminating the class imbalance, however, they do not handle the complexities of sequential data. In this study, using the data of over 30,000 students from the Open University (UK), we implement a deep-learning-based approach using adversarial networks, Sequential Conditional Generative Adversarial Network (SC-GAN) that encapsulates the past behavior of each student for its previous sequences and generates synthetic student records for the next timestamp. The proposed approach is devised to generate instances, which are augmented with the actual data to eliminate class imbalance. A performance comparison of the proposed SC-GAN with the standard up-sampling methods is also presented and the results validate the proposed method with an improved AUC of 7.07% and 6.53%, respectively, when compared with conventional Random Over-sampling and Sythetic Minority Oversampling techniques.