Association Networks in Time Series Data Mining

A new method of time series data mining using moving approximation (MAP) transform and association measures based on MAP is discussed and an effective method of MAP transform calculation for time series with fixed time step is proposed.

Abstract

We discuss a new method of time series data mining using moving approximation (MAP) transform and association measures based on MAP. MAP transform replaces time series values by slope values of lines approximating time series data in sliding window. An effective method of MAP transform calculation for time series with fixed time step is proposed. Based on MAP, a measure of local trend associations between time series is introduced. This measure is invariant under independent linear transformations of time series. Measure of local trend associations defines association function depending on the size of sliding window for each pare of considered time series. Based on association function, different association measures may be considered to measure local trend associations or global trend associations between time series. The methods of application of association measure to construction of association network of time series are discussed and illustrated on examples of synthetic and financial time series databases. Association networks give information about relationships between time dynamics of elements of systems given by time series databases.