Improved Models for Time Series Cluster Representation Based Dynamic Time Warping

Abstract

Revolution of Smart Things (SThs) connected to the Internet to build Internet of Things (IoT) applications, causes a flood of data streams every moment. Main root causes of massive SThs integration for increasing accuracy of sensed features and for enabling fault tolerance. In general, resulting deluge of real-time data streams has the property of five V of the big data (i.e., volume, velocity, variety, veracity, and value). Such properties make mining and analysis of massive and heterogeneous data be challenging tasks. In our previous work, we present three novel data reduction models based on Dynamic Time Warping (DTW) for enabling balanced indexing in the IoT. This paper presents two extensions to improve the Hybrid algorithm (ClRe 3.0) using DTW warped path. First extension (ClRe 3.1): targets improving accuracy of indexed clusters representatives by taking the average of individual warped items and keeping only 50% of the warped items for each warped slot. Second extension (ClRe 3.2): targets decreasing size of indexed clusters representatives as possible by compensating every warped slot by its corresponding item keeping only common items with minimum distances. The proposed extensions are explained using real samples and evaluated using Szeged-weather dataset as well. The evaluation results proves that ClRe 3.1 could enhance the accuracy of ClRe 3.0 by approximate 9% in average, keeping indexes sizes as possible as fitted (i.e., < the average length of all datasets). In case of indexing only highly common readings, ClRe 3.2 out-performs other extensions in decreasing indexes sizes.