No abstract

Abstract. This paper introduces a new method for explaining the predictions of ensembles of neural networks on a case by case basis. The approach of explaining individual examples differs from much of the current research which focuses on producing a global model of the phenomenon under investigation. Explaining individual results is accomplished by modelling each of the networks as a rule-set and computing the resulting coverage statistics for each rule given the data used to train the network. This coverage information is then used to choose the rule or rules that best describe the example under investigation. This approach is based on the premise that ensembles perform an implicit problem space decomposition with ensemble members specialising in different regions of the problem space. Thus explaining an ensemble involves explaining the ensemble members that best fit the example.