A Parallel Approach to Discords Discovery in Massive Time Series Data

A discord is a refinement of the concept of an anomalous subsequence of a time series. Being one of the topical issues of time series mining, discords discovery is applied in a wide range of real-world areas (medicine, astronomy, economics, climate modeling, predictive maintenance, energy consumption, etc.). In this article, we propose a novel parallel algorithm for discords discovery on high-performance cluster with nodes based on many-core accelerators in the case when time series cannot fit in the main memory. We assumed that the time series is partitioned across the cluster nodes and achieved parallelization among the cluster nodes as well as within a single node. Within a cluster node, the algorithm employs a set of matrix data structures to store and index the subsequences of a time series, and to provide an efficient vectorization of computations on the accelerator. At each node, the algorithm processes its own partition and performs in two phases, namely candidate selection and discord refinement, with each phase requiring one linear scan through the partition. Then the local discords found are combined into the global candidate set and transmitted to each cluster node. Next, a node performs refinement of the global candidate set over its own partition resulting in the local true discord set. Finally, the global true discords set is constructed as intersection of the local true discord sets. The experimental evaluation on the real computer cluster with real and synthetic time series shows a high scalability of the proposed algorithm.