Discovering Triggering Events from Longitudinal Data

Abstract

Longitudinal data consist of the repeated measurements of some variables which describe the dynamics of a domain(process or phenomenon) over time. They can be analyzed in order to explain what event may cause the transition from a state into the next one during the evolution of the domain. Generally, approaches to this explanation problem rely on the exclusive usage of domain knowledge, while an analysis driven from only data is still lacking. In this paper we describe a data mining approach to discover events which may have triggered a transition during the evolution of the domain. The original data mining task is decomposed into two consecutive subtasks. First, the sequence of discrete states which represents the dynamics of the domain is determined. Second, the triggering events for two successive states are found out. Computational solutions to both problems are presented. Their application to two real scenarios is presented and results are discussed.