Outlier Explanation Through Masking Models

Given a database and one single anomalous data point, the Outlying Aspect Mining problem consists in explaining the abnormality of that data point w.r.t. the data population stored in the input database. Thus, the problem requires the discovery of the sets of attributes and associated values that account for the abnormality of a data point within a given data set. In this setting, the abnormality of the data point at hand is stated beforehand, e.g., as the result of some outlier detection techniques (which, for the most part, do not provide information about why the selected data points are actually anomalous). This paper proposes a solution to the OAM problem exploiting a deep learning architecture. Besides explaining the input data point abnormality by singling out the smallest set of pairs attribute-value justifying it, our technique also provides new values for those attributes that would transform the input outlier into an inlier. Several experiments are also presented that assess the effectiveness of our approach.