An Approach for Discovery of Complex Events and Hierarchies

Abstract

This paper presents techniques for discovery of event hierarchies in event streams. Event discovery is about recognition of low level events; their relationships and how they combine to cause a composite event. The challenge is that the occurrence time of a composite event, the identity of the low level events, the number of the low level events, and relationship are not known in advance. We start by identifying a set of ‘candidate events’ that lead to a given composite event. We then filter the candidate events to establish the actual events that lead to the composite event. Then a causal relationship between the filtered low level events is discovered. The causal relation among the events allows generation of a hierarchical structure that shows the composition structure between low level events and the composite event. We rely on domain experts and literature to identify the initial set of low level candidate events. To filter candidate events, we use a historical event stream. We develop an approach based on heuristics and similarity measures to identify the structural relations between low level events and composite event. The discovered structure of the events is then validated using domain experts. The approach was developed using a case study of financial crisis with the historical news corpus archived by major news networks, particularly CNN as the event stream.