Pitfalls in benchmarking data stream classification and how to avoid them

摘要: Data stream classification plays an important role in modern data analysis, where arrives a and needs to be mined real time. In the setting underlying distribution from which this comes may changing evolving, so classifiers that can update themselves during operation are becoming state-of-the-art. paper we show streams have temporal component, currently is not considered evaluation benchmarking of classifiers. We demonstrate how naive classifier considering component only outperforms lot current state-of-the-art on dependence, i.e. autocorrelated. propose evaluate taking into account introduce new measure, provides more accurate gauge performance. response dependence issue generic wrapper for classifiers, incorporates attribute space.