A Classification Tree Method Based on Belief Entropy for Evidential Data

Decision tree is widely applied in classification and recognition areas, but meanwhile it is hard to learn from evidential data with uncertainty. To solve this issue, we propose a decision tree method which can learn from uncertain data sets and guarantee a certain classification performance when handle problems with huge ignorance or uncertainty. This tree method selects attribute based on belief entropy, a kind of uncertainty measurement, which is calculated from the basic belief assignment. And especially the Evidential Expectation-Maximization algorithm is adopted to extract the distribution parameters from evidential likelihood to generate the basic belief assignment. The proposed method is an extension of decision tree based on belief entropy, which is supposed to handle problems with precise data. Some numerical experiments on Iris and Sonar data set are conducted and the experimental results suggested that the proposed tree method achieves good result on data with high-level uncertainty.