A Framework for Classification in Data Streams Using Multi-strategy Learning

Abstract

Adaptive online learning algorithms have been successfully applied to fast-evolving data streams. Such streams are susceptible to concept drift, which implies that the most suitable type of classifier often changes over time. In this setting, a system that is able to seamlessly select the type of learner that presents the current “best” model holds much value. For example, in a scenario such as user profiling for security applications, model adaptation is of the utmost importance. We have implemented a multi-strategy framework, the so-called Tornado environment, which is able to run multiple and diverse classifiers simultaneously for decision making. In our framework, the current learner with the highest performance, at a specific point in time, is selected and the corresponding model is then provided to the user. In our implementation, we employ an Error-Memory-Runtime (EMR) measure which combines the error-rate, the memory usage and the runtime of classifiers as a performance indicator. We conducted experiments on synthetic and real-world datasets with the Hoeffding Tree, Naive Bayes, Perceptron, K-Nearest Neighbours and Decision Stumps algorithms. Our results indicate that our environment is able to adapt to changes and to continuously select the best current type of classifier, as the data evolve.