Tapped Delay Lines for GP Streaming Data Classification with Label Budgets

Abstract

Streaming data classification requires that a model be available for classifying stream content while simultaneously detecting and reacting to changes to the underlying process generating the data. Given that only a fraction of the stream is ‘visible’ at any point in time (i.e. some form of window interface) then it is difficult to place any guarantee on a classifier encountering a ‘well mixed’ distribution of classes across the stream. Moreover, streaming data classifiers are also required to operate under a limited label budget (labelling all the data is too expensive). We take these requirements to motivate the use of an active learning strategy for decoupling genetic programming training epochs from stream throughput. The content of a data subset is controlled by a combination of Pareto archiving and stochastic sampling. In addition, a significant benefit is attributed to support for a tapped delay line (TDL) interface to the stream, but this also increases the dimensionality of the task. We demonstrate that the benefits of assuming the TDL can be maintained through the use of oversampling without recourse to additional label information. Benchmarking on 4 dataset demonstrates that the approach is particularly effective when reacting to shifts in the underlying properties of the stream. Moreover, an online formulation for class-wise detection rate is assumed, where this is able to robustly characterize classifier performance throughout the stream.