Semi-Supervised Self-training Approaches for Imbalanced Splice Site Datasets

Machine Learning algorithms produce accurate classifiers when trained on large, balanced datasets. However, it is generally expensive to acquire labeled data, while unlabeled data is available in much larger amounts. A cost-effective alternative is to use Semi-Supervised Learning, which uses unlabeled data to improve supervised classifiers. Furthermore, for many practical problems, data often exhibits imbalanced class distributions and learning becomes more challenging for both supervised and semisupervised learning scenarios. While the problem of supervised learning from imbalanced data has been extensively studied, it has not been studied much for semi-supervised learning. Thus, in this study, we carry out an empirical evaluation of a semi-supervised learning algorithm, specifically self-training based on Näıve Bayes Multinomial (NBM), and address the issue of imbalanced class distributions both at data-level (by re-sampling) and algorithmic-level (using cost-sensitive learning and ensembles). We conduct our study on the problem of splice site prediction, a problem for which the ratio of positive to negative examples is very high. Our experiments on five different datasets show that a simple method that adds only positive instances to the labeled data in the semi-supervised iterations produces consistently better results when compared with other methods that deal with data imbalance.