Abstract/Summary

Data streams are unbounded, sequential data instances that are generated with high Velocity. Data streams arrive online (i.e., instance by instance) and there is no control over the order in which data instances arrive either within a data stream or across data streams. Classifying sequential data instances is a challenging problem in machine learning with applications in network intrusion detection, financial markets and sensor networks. The automatic labelling of unseen instances from the stream in real-time is the main challenge that data stream classification faces. For this, the classifier needs to adapt to concept drifts and can only have a single-pass through the data with a limited amount of memory if the stream is generating data instances at a high Velocity. Nowadays the focus of Data Stream Mining (DSM) lies in the development of data mining algorithms rather than on pre-processing techniques. To the best of the author knowledge, at present, there are no developments for truly real-time feature selection in a streaming setting. This research work presents a real-time pre-processing technique, in particular, feature tracking in combination with concept drift detection. The feature tracking is designed to improve DSM classification algorithms by enabling real-time feature selection. The pre-processing technique is based on tracking adaptive statistical summaries of the data and class label distributions, known as Micro-Clusters. Thus the three objectives of this research were to develop a real-time pre-processing technique that can (1) detect a concept drift, (2) identify features that were involved in concept drift and thus potentially change their relevance and (3) build a real-time feature selection method based on the developments mentioned above. The evaluation of the developed technique is based on artificial data streams with known ground truth and real datasets with and without artificially induced concept drift (i.e., controlled and uncontrolled real datasets). It was observed that the developed method for concept drift detection did detect induced concept drifts very well compared with alternative concept drift detection methods. Overall the research represents a first attempt to resolve real-time feature selection for DSM tasks. It has been shown that the technique can indeed identify concept drift, track features, and identify features that may have changed their relevance for the DSM task in real-time. It has also been shown that the developed method for real-time feature selection can improve the accuracy of data stream classification tasks.