Refined classifier combination using belief functions

We address here the problem of supervised classification using belief functions. In particular, we study the combination of non-independent sources of information. In a companion paper, we showed that the cautious rule of combination may be best suited than the widely used Dempsterpsilas Rule to combine classifiers in the case of real data. Then, we considered combination rules intermediate between the cautious rule and Dempsterpsilas rule. We proposed a method for choosing the combination rule that optimizes the classification accuracy over a set of data. Eventually, we mentioned a generalized approach, in which a refined combination rule best suited to complex dependencies of the sources to combine is learnt. Here, we extensively study this latter approach. It consists in clustering the sources according to some measure of similarity; then, one rule is learnt for combining the sources within the clusters, and another for combining the results thus obtained. We conduct experiments on various real data sets that show the interest of this approach.