Adversarial Anomaly Explanation

Given a data set and one single object known to be anomalous beforehand, the outlier explanation problem consists in explaining the abnormality of the input object with respect to the data set population. The approach pursued in this paper to solve the above task consists in finding an explanation, namely, a piece of information encoding the characteristics that locate the anomalous data object far from the normal data. Our explanation consists of two components, the choice, encoding the set of features in which the anomalous object deviates from the rest of the population, and the mask, encoding the associated amount of deviation with respect to the normality. The goal here is not to explain the decisional process of a model but, rather, to provide an explanation justifying the output of the decisional process by only inspecting the data set on which the decision has been made. We tackle this problem by introducing an innovative deep learning architecture, called MMOAM, based on the adversarial learning paradigm. We assess the effectiveness of our technique over both synthetic and real data sets and compare it against state of the art outlier explanation methods reporting better performances in different scenarios.