Bernhard Pfahringer discusses how to deal with change in data mining, and introduces ADWIN, an adaptive windowing method that grows and shrinks.
Share this step
Change is everywhere! – and is a distinguishing feature of data stream mining. Bernhard Pfahringer explains that one way of dealing with change is to use an adaptive windowing method called ADWIN that grows a sliding window on the data stream in times of stability and shrinks it in times of change. The Hoeffding Adaptive Tree grows alternative branches and monitors their performance using ADWIN. Bagging, which involves bootstrap sampling with replacement, can be turned into an online algorithm using a weighting technique, and coupled with ADWIN for explicit change detection. MOA’s data stream generators can simulate change over time, allowing these mining algorithms to be tested and compared.