B1819
Title: An anticipative Bayesian stream classifier for data streams with verification latency
Authors: Georg Krempl - University of Utrecht (Netherlands)
Vera Hofer - University of Graz (Austria) [presenting]
Dominik Lang - (Germany)
Abstract: One of the challenges of classification in non-stationary data streams is updating the classification rule after distributional changes in the case of verification latency. Various existing techniques assume at least a small number of recent labelled data. Such recent data is often missing under verification latency. An anticipative Bayesian stream classifier (ABClass) is proposed for such a situation. ABClass uses density estimation techniques, extended to extrapolate drift patterns over time. It applies unsupervised parameter tuning and unsupervised model selection. ABClass is generic, which allows the inclusion of different types of drift models, both for the class-conditional feature distribution as well as for the class-prior distribution. The various classification techniques among which ABClass can select the most appropriate one can easily be extended to make ABClass highly flexible and adaptive. The performance of ABClass is evaluated in experiments based on real-world data streams. The results are compared to other state-of-the-art approaches.
