Concept Drift Detection Using Online Histogram-Based Bayesian Classifiers

Abstract

In this paper, we present a novel algorithm that performs online histogram-based classification, i.e., specifically designed for the case when the data is dynamic and its distribution is non-stationary. Our method, called the Online Histogram-based Naïve Bayes Classifier (OHNBC) involves a statistical classifier based on the well-established Bayesian theory, but which makes some assumptions with respect to the independence of the attributes. Moreover, this classifier generates a prediction model using uni-dimensional histograms, whose segments or buckets are fixed in terms of their cardinalities but dynamic in terms of their widths. Additionally, our algorithm invokes the principles of information theory to automatically identify changes in the performance of the classifier, and consequently, forces the reconstruction of the classification model in run-time as and when it is needed. These properties have been confirmed experimentally over numerous data sets (In the interest of space and brevity, we present here only a subset of the available results. More detailed results are found in [2].) from different domains. As far as we know, our histogram-based Naïve Bayes classification paradigm for time-varying datasets is both novel and of a pioneering sort.

C.A. Astudillo—This work was partially supported by the FONDECYT Grant No. 11121350, Chile.

B.J. Oommen—Chancellor’s Professor; Fellow: IEEE and Fellow: IAPR. This author is also an Adjunct Professor with the University of Agder in Grimstad, Norway. The work of this author was partially supported by NSERC, the Natural Sciences and Engineering Research Council of Canada.