Cluster-Based Similarity Search in Time Series

In this paper, we present a new method that accelerates similarity search implemented via one-nearest neighbor on time series data. The main idea is to identify the most similar time series to a given query without necessarily searching over the whole database. Our method is based on partitioning the search space by applying the K-means algorithm on the data. Then, similarity search is performed hierarchically starting from the cluster that lies most closely to the query. This procedure aims at reaching the most similar series without searching all clusters. In this work, we propose to reduce the intrinsically high dimensionality of time series prior to clustering by applying a well known dimensionality reduction technique, namely, the Piecewise Aggregate Approximation, for its simplicity and efficiency. Experiments are conducted on twelve real-world and synthetic datasets covering a wide range of applications.