Visual analytics e outlying aspect mining: contextualização de anomalias considerando questões temporais e multidimensionais

Abstract

Outlying Aspect Mining (OAM) is a new way of handling outliers that, instead of focusing solely on the detection, also provides an explanation for the abnormal status. For this purpose, a subspace of attributes considered as the most relevant for understanding the sample outlying aspects is presented. There are many challenges associated with the application of OAM, such as combinatorial explosion of the search space and ability to compare metrics calculated for subspaces with different dimensionalities. Even só, listing a group of attributes is not sufficient for a human specialist to comprehend the situation and take the necessary actions. A higher-level, visual approach can improve the process by providing better cognitive clues to experts. Here we describe the application of an OAM technique in a fault detection problem for locomotives. Based on the experience obtained in this use case, we proposed and developed a Visual Analytics platform for the processing and representation of data in a user-friendly interface. A novelty available on this platform are parallel coordinates plots that also display temporal multidimensional data. Such representation tries to circumvent human visual system limitations and helps the outlier investigation. To explore and validate the applicability of the developed tool, the locomotive operation use case is employed again.