Data Abstraction for Visualizing Large Time Series

Abstract

Numeric time series is a class of data consisting of chronologically ordered observations represented by numeric values. Much of the data in various domains, such as financial, medical and scientific, are represented in the form of time series. To cope with the increasing sizes of datasets, numerous approaches for abstracting large temporal data are developed in the area of data mining. Many of them proved to be useful for time series visualization. However, despite the existence of numerous surveys on time series mining and visualization, there is no comprehensive classification of the existing methods based on the needs of visualization designers. We propose a classification framework that defines essential criteria for selecting an abstraction method with an eye to subsequent visualization and support of users' analysis tasks. We show that approaches developed in the data mining field are capable of creating representations that are useful for visualizing time series data. We evaluate these methods in terms of the defined criteria and provide a summary table that can be easily used for selecting suitable abstraction methods depending on data properties, desirable form of representation, behaviour features to be studied, required accuracy and level of detail, and the necessity of efficient search and querying. We also indicate directions for possible extension of the proposed classification framework.

Visual‐Interactive Preprocessing of Multivariate Time Series Data

“…In the data mining, machine learning, and information retrieval, these downstream steps include content‐based information retrieval [LSDJ06], indexing [Mül07], tracking [MHK06], similarity search [KTWZ10], feature analysis [Mör06], descriptor analysis [KK03], motif discovery [Fu11], anomaly detection [SMF15], rule discovery [Mor06], classification [MR06], clustering [WL05] segmentation [BBB∗18], labeling [BDV∗17], prediction [EA12], monitoring [LKL∗04], or exploratory search [Ber15]. In visualization approaches these downstream steps are additionally conflated with the goal to support users with effective visualizations [AMST11] requiring meaningful data preparation [SAAF18].…”

“…Most data reduction techniques eliminate irrelevant parts of the data while preserving relevant information. Compact data representations help to improve the performance and scalability of the subsequent analysis steps as well as visualization approaches [SAAF18]. Important classes of techniques include sampling and filtering , with the overall idea to reduce information with respect to a pre‐defined criterion [Fu11], As an alternative, descriptors can be applied to revive compact representations of the data [KK03, EA12].…”

“…In the data mining, machine learning, and information retrieval, these downstream steps include content-based information retrieval [LSDJ06] [Ber15]. In visualization approaches these downstream steps are additionally conflated with the goal to support users with effective visualizations [AMST11] requiring meaningful data preparation [SAAF18].…”

COPE: Interactive Exploration of Co-Occurrence Patterns in Spatial Time Series

“…Data noise. In time series analysis, it is acknowledged that data may have irregular fluctuations, that is, noise is treated as an indispensable component of time series [47]. Noise and missing values can be handled in our system using techniques for data smoothing and interpolation over a sliding temporal window.…”

Data Abstraction and Pattern Identification in Time-series Data

“…Data abstraction can transform the raw data samples into a simple format while at the same time preserving significant features that are important for the user [31,110,118]. Large data sets containing billions of raw data samples can be pre-processed using standard data abstraction techniques and can be converted into a simple format.…”