Multi-represented classification based on confidence estimation

Abstract. Complex objects are often described by multiple representations mod-eling various aspects and using various feature transformations. To integrate all information into classification, the common way is to train a classifier on each representation and combine the results based on the local class probabilities. In this paper, we derive so-called confidence estimates for each of the classifiers reflecting the correctness of the local class prediction and use the prediction hav-ing the maximum confidence value. The confidence estimates are based on the distance to the class border and can be derived for various types of classifiers like support vector machines, k-nearest neighbor classifiers, Bayes classifiers, and decision trees. In our experimental results, we report encouraging results demonstrating a performance advantage of our new multi-represented classifier compared to standard methods based on confidence vectors.