Finding similar time series

Abstract

Similarity of objects is one of the crucial concepts in several applications, including data mining. For complex objects, similarity is nontrivial to define. In this paper we present an intuitive model for measuring the similarity between two time series. The model takes into account outliers, different scaling functions, and variable sampling rates. Using methods from computational geometry, we show that this notion of similarity can be computed in polynomial time. Using statistical approximation techniques, the algorithms can be speeded up considerably. We give preliminary experimental results that show the naturalness of the notion.