6533b861fe1ef96bd12c5a72

RESEARCH PRODUCT

Discovering representative models in large time series databases

subject

description

The discovery of frequently occurring patterns in a time series could be important in several application contexts. As an example, the analysis of frequent patterns in biomedical observations could allow to perform diagnosis and/or prognosis. Moreover, the efficient discovery of frequent patterns may play an important role in several data mining tasks such as association rule discovery, clustering and classification. However, in order to identify interesting repetitions, it is necessary to allow errors in the matching patterns; in this context, it is difficult to select one pattern particularly suited to represent the set of similar ones, whereas modelling this set with a single model could be more effective. In this paper we present an approach for deriving representative models in a time series. Each model represents a set of similar patterns in the time series. The approach presents the following peculiarities: (i) it works on discretized time series but its complexity does not depend on the cardinality of the alphabet exploited for the discretization; (ii) derived models allow to express the distribution of the represented patterns; (iii) all interesting models are derived in a single scan of the time series. The paper reports the results of some experimental tests and compares the proposed approach with related ones.