Tell Me Why - A Systematic Literature Review on Outlier Explanation for Tabular Data

Outlier explanation approaches are used to support analysts in investigating outliers, especially those detected by methods which are not intuitively interpretable such as deep learning or ensemble approaches. There have been several studies on outlier explanation and visualization in the last years. Nonetheless, there has been no overarching survey of existing outlier explanation approaches. We investigate what kind of explanation approaches exist for tabular data and how they can be characterized. For this, we perform a literature review of 418 papers uncovering 26 existing studies which we analyze and categorize according to: approach, model dependence, supported data, output and evaluation. Based on our categorization, we identify trends within outlier explanation. This includes the limited number of underlying approaches and the lack of a common evaluation method. We are also able to highlight avenues for further research such as utilizing different datasets for evaluation or exploring visualization approaches.