Quality of classification explanations with PRBF

Recently two general methods for explaining classification models and their predictions have been introduced. Both methods are based on an idea that importance of a feature or a group of features in a specific model can be estimated by simulating lack of knowledge about the values of the feature(s). For the majority of models this requires an approximation by averaging over all possible feature values. A probabilistic radial basis function network (PRBF) is one of the models where such approximation is not necessary and therefore offers a chance to evaluate the quality of approximation by comparing it to the exact solution. We present both explanation methods and demonstrate their behavior with PRBF. The explanations make individual decisions of classifiers transparent and allow inspection and visualization of otherwise opaque models. We empirically compare the quality of explanations based on marginalization of the Gaussian distribution (the exact method) and explanation with averaging over all feature values (the approximation). The results show that the approximation method and the exact solution give very similar results, which increases the confidence in the explanation methodology also for other classification models.