Classification of Multivariate Time Series via Temporal Abstraction

Analysis of multivariate time stamped data, for purposes such as Temporal Knowledge Discovery, Classification and Clustering, presents many challenges. Time stamped data can be sampled in a fixed frequency, commonly when measured by electronic sensors, but also in a non fixed frequency, often when made manually. Additionally, raw temporal data can represent periods of a continuous or nominal value represented by time intervals. Temporal bstraction, in which time point series are abstracted into meaningful time intervals, is used to bring all the temporal variables to the same representation. In this talk I will present KarmaLego, a fast time intervals mining method for the discovery of non-ambiguous Time Intervals Related Patterns (TIRP) represented by Allen's temporal relations. Then I will present several uses of the TIRPs. In this talk I will focus on the use of classification of multivariate time series, in which TIRPs are used as classification features. The entire process and several computational abstraction methods for the task for classification will be presented. Finally an evaluation on real datasets from the medical domain will be presented, in addition to meaning examples of discovered temporal knowledge.