Outlier detection provides the methodology to identify issues with data quality, to spot unusual data items or to detect rare patterns in data. Interpreting these outliers still remains a challenging task. Data analysts or domain experts trying to understand the nature of outliers cannot be expected to compare outliers to the remaining data to understand the differences. Outlier description therefore aims to provide additional information that supports interpretation by characterizing the differences.
In this talk, I discuss our recent work in outlier detection and description, and highlight some of the current challenges that I consider to be of prime importance for mature outlier description.

Outlier detection has been studied extensively and employed in diverse applications in the past decades. Yet the related and important outlier description problem has been an understudied problem. This problem often arises in practice when we have a small number of data instances that had been identified to be outliers and we wish to explain why they are outliers. Consider the example of automobile recalls where we already have examples of the outliers (recalled cars) and many examples of the non-outliers (non-recalled cars), with the aim to find a description of what made the instances anomalous.
We formalize the outlier description problem, present complexity results and declarative formulations which offers great flexibility in incorporating diverse scenarios arising in practice such as multiple explanations and human in the loop extensions.

In this talk I will revisit the challenge of outlier description and interpretation. Essentially, can we manifest why an object is an outlier? We consider several possible ways to produce description and interpretation, including at the data level, at the feature/subspace, and at the model level. We discuss some example applications of such description and interpretation and examine their pros and cons. The talk will not come to a decisive conclusion about techniques. Instead, hopefully it will inspire a few challenges for future work.

Outlier detection has historically been used in a variety of domains for identifying rare samples in datasets. In this talk, I will discuss the particular application of outlier detection for identifying misbehavior and malicious actors in online social platforms. I will overview several previous works which leverage outlier detection techniques for identifying abusive following behavior and fake viewership on online platforms, and discuss technical takeaways and meta-lessons learned from these projects and more. The purpose of this talk is to communicate the great value that outlier detection approaches add in misbehavior detection, but also to pose a call to action for prioritizing the discovery or identification of specific outlying behaviors as opposed to outliers in general.