Exploiting efficient parallelism for mining rules in time series data

Abstract

Mining interesting rules from time series data has earned a lot of attention to the data mining community recently. It is quite useful to extract important patterns from time series data to understand how the current and the past values of patterns in the multivariate time series data are related to the future. These relations can basically be expressed as rules. Mining these interesting rules among patterns is time consuming and expensive in multi-stream data. Incorporating parallel processing techniques is helpful to solve the problem. In this paper, we present a parallel algorithm based on a lattice theoretic approach to find out the rules among patterns that sustain sequential nature in the multi-stream data of time series. The human motion data considered as multi-stream multidimensional data used as data set for this purpose is transformed into sequences of symbols of lower dimension due to its complex nature. Then the proposed algorithm is implemented on a Distributed Shared Memory (DSM) multiprocessors system. The experimental results justify the efficiency of finding rules from the sequences of the patterns for time series data by achieving significant speed up comparing with the previous reported algorithm. © Springer-Verlag Berlin Heidelberg 2005.