EXOS: Explaining Outliers in Data Streams

Abstract

Real-time outlier detection is important in many data stream applications. To help analysts understand the detected outliers better, the outliers should be presented with their explanations. One type of explanations for an outlier is its set of outlying attributes which is a subset of features responsible for the outlier’s abnormality. There exist techniques that generate outlying attributes in data streams; however, none simultaneously considers the cross-correlation among data streams, the unbounded volume of data, and concept drift. To fill this gap, we propose EXOS, a framework that generates outlying attributes in multi-dimensional data streams. For each outlier, it incrementally finds a local context to determine the decision boundary that separates the outlier from the normal data while handling both the unbounded volume of data and concept drift. It considers the potential data correlation within a data stream and across data streams to estimate the local context. The experiments using three real and two synthetic datasets show that, on average, EXOS achieves up to 49% higher F1 score and 29.6 times lower explanation time than existing algorithms.