A Metaheuristic for Classification of Interval Data in Changing Environments

The Bayes approach is arguably the classification method most used in unspecialized applications, thanks to its robustness, simplicity, and interpretability. The main problem here is establishing proper probability values. This paper deals with adapting the above method for cases where the classified data is of interval type, with changing environments (evolving data stream, concept drift, nonstationarity). The probability values are estimated using nonparametric methods, thanks to which the procedure becomes independent of characteristics of learning subsets representing particular classes. They can also be supplemented with new, current observations, added while performing the algorithm. The investigated process also removes elements with negligible or even negative impact on accuracy of results, which increases the effectiveness of adaptation in conditions of changing reality. It is possible to differentiate the meanings of particular classes. The method allows any number of them. The particular attributes of data elements may be continuous, categorical, or both.