Evidence combination in medical data mining

In this work we apply Dempster-Shafer´s theory of evidence combination for mining medical data. We consider the classification task in two domains: breast tumors and skin lesions. Classifier outputs are used as a basis for computing beliefs. Dynamic uncertainty assessment is based on class differentiation. We combine the beliefs of three classifiers: k-nearest neighbor (kNN), naive Bayesian and decision tree. Dempster´s rule of combination combines three beliefs to arrive at one final decision. Our experiments with k-fold cross validation show that the nature of the data set has a bigger impact on some classifiers than others and the classification based on combined belief shows better overall accuracy than any individual classifier. We compare the performance of Dempster´s combination (with differentiation-based uncertainty assignment) with those of performance-based linear and majority vote combination models. We study the circumstances under which the evidence combination approach improves classification.