A Context-Sensitive Framework for Mining Concept Drifting Data Streams

In this chapter, we present the staged learning approach to classification in a non-stationary stream of data. Unlike the standard data stream mining paradigm that assumes change is always present, the staged approach senses the level of volatility in the stream and adjusts the mode of learning accordingly. We propose a scheme whereby volatility could be measured and construct a volatility detector that senses the stream. We model the data stream as consisting of two states: a high-volatility state and a low-volatility state, with transitions taking place to/from these states depending on the level of volatility in the stream. In segments of high volatility an ensemble of online classifiers is used for learning, whereas in low volatility maximum utilization is made of past concepts which are encoded by compact versions of Fourier spectra. The staged approach results in improvements in accuracy as well as throughput while reducing memory usage as demonstrated by our experimentation on a wide range of real-world and synthetic datasets.