Adapting a Combination Rule to Non-Independent Information

B. Quost, T. Denoeux, M.-H. Masson

In this article, we address the combination of non-independent sources to solve classification problems, within the theory of belief functions. We show that the cautious rule of combination [1, 2] is well-suited to such problems. We propose a method to learn the combination rule from training data, and we generalize it in the case of complex dependence of the sources. We demonstrate the validity of our approach on several synthetic and realdata sets.

PDF full paper