Anomaly Detection via Correlation Clustering

Anomaly detection has been studied and modeled using various methods, mostly statistical but also combinatorial, including clustering. This paper investigates the utility of using exact correlation clustering modeled via the Cluster Editing approach for accurate anomaly detection. We use Twitter data for this purpose. The data pipeline produces a sequence of graphs from text to represent relations between entities (in this case, tweets.) Each graph constitutes a single data element, thus a vertex in a final graph on which the clustering is performed.