Identification of Outliers in Graph Signals
Karthik Gopalakrishnan, Max Z. Li, Hamsa Balakrishnan · 2019
Outlier detection, or the identification of observations that differ significantly from the norm, is an important aspect of data mining. Conventional outlier detection tools have limited applicability to networks, in which there are interdependencies between the variables. In this paper, we consider the problem of identifying unusual spatial distributions of nodal signals on a graph. Leveraging tools from graph signal processing and statistical analysis, we propose a methodology to identify outliers in graph signals in a computationally efficient manner. Specifically, we examine a projection of the graph signal into a lower dimensional representation that enables easier outlier identification. Additionally, we derive analytical expressions for the outlier bounds. We apply our technique by identifying off-nominal days in the context of the US airport network using aviation delay data.