Balancing Between Scalability and Accuracy in Time-Series Classification for Stream and Batch Settings

Abstract

As big data sources providing time series increase, and data is provided in increased velocity and volume, we need to efficiently recognize data provided, classifying it according to their type, origin etc. This is a first important step in doing analytics on data provided from disparate data sources, such as archival sources, multiple sensors, or social media feeds. Time series classification is the task labeling time series using a set of predefined labels. In this paper we present the K-BOSS-VS algorithm for time series classification. The proposed algorithm is based on state-of-the-art symbolic time series classification algorithms, and aims to achieve high accuracy, balancing with computational efficiency. K-BOSS-VS exploits K representatives of each time series class to classify new series. This provides opportunities for representing intra-class differences, thus increasing the classification accuracy, while incurring a small performance overhead compared to methods using one class representative. Additionally, K-BOSS-VS offers a solution for classifying time-series in batch and streaming settings, due to the opportunities for increasing computational efficiency and the low memory requirements.