Experimenting with Supervised Drift Detectors in Semi-supervised Learning

Machine learning algorithms to aid decision-making processes are increasingly common in several areas. When fully-trained, algorithms tend to perform better, but the availability of data labels shortly after testing without human intervention is a challenging task in many areas, especially in data stream learning with concept drifts, where data is generated very fast, in real-time, with the possibility of changes in the data distribution. Concept drifts have been addressed in different ways, but using drift detectors with base classifiers in semi-supervised learning is not so common. This article shows how to use state-of-the-art supervised detectors in semi-supervised learning problems, and it also includes an extension to the MOA framework. The Experiments designed to test our proposal used Hoeffding Tree (HT) as base classifier, combined with eight drift detectors and a total of 62 artificial and five real-world datasets, configured with 15 % and 30 % labeled instances. The results indicate that drift detectors designed for supervised learning can also be effectively used in semi-supervised environments. This finding could lead to a change of paradigm for future research, since supervised drift detectors have never been considered as a viable alternative due to the absence of labels shortly after testing in many real-world data streams.