Abstract

In  multidimensional  classification  the  goal  is  to assign  an  instance  to  a  set  of  different  classes.This  task  is  normally  addressed  either  by  defin-ing a compound class variable with all the possi-ble combinations of classes (label power-set meth-ods, LPMs) or by building independent classifiers for each class (binary-relevance methods, BRMs).However,  LPMs do not scale well and BRMs ig-nore the dependency relations between classes. We introduce  a  method  for  chaining  binary  Bayesian classifiers  that  combines  the  strengths  of  classi-fier chains and Bayesian networks for multidimen-sional classification.   The method consists of two phases. In the first phase, a Bayesian network (BN)that  represents  the  dependency  relations  between the class variables is learned from data. In the sec-ond phase, several chain classifiers are built, such that  the  order  of  the  class  variables  in  the  chain is  consistent  with  the  class  BN.  At  the  end  we combine the results of the different generated or-ders.  Our method considers the dependencies be-tween  class  variables  and  takes  advantage  of  the conditional independence relations to build simpli-fied models. We perform experiments with a chain of na ıve Bayes classifiers on different benchmark multidimensional  datasets  and  show  that  our  ap-proach outperforms other state-of-the-art methods