Rapidly Labeling and Tracking Dynamically Evolving Concepts in Data Streams

Abstract

Data mining research has produced a significant repertoire of algorithms to predict the classification of data instances with reasonable accuracy. However, data quantity and availability is continuing to rapidly expand such that we no longer have fixed and manageable data sets, but rather continual streams of data. Mining streaming data becomes challenging when using a piece-wise or online approach, however, due to concept drift and feature evolution. As the stream progresses, features may be added, removed, or change in the range of possible values, which is known as feature evolution. The defining concepts for a label or class may also migrate over the span of the data stream. Novel classes that were not known a priori can also appear within the data stream. Traditional algorithms often characterize unknown labels as errors and outliers, but in the dynamic streaming domain, a sufficiently dense cluster of outliers must be analyzed to discover potential emergent classes. In our approach, we aim to adapt to novel classes emergence, feature evolution, and concept drift while labeling data instances and adhering to the constraints of continuous data streams. Our solution consists of an adaptive supervised ensemble for predicting instance labels, and a stream clustering approach to monitor concept defining characteristics and novel class development without regard to label (i.e. unsupervised). We analyze and compare our approach to traditional baseline approaches on benchmark data streams to verify the accuracy and efficiency of the algorithm.