Interactive Visualization and Event Detection in Time-series Data
Visualization enriches human understanding of data through visual representations across many levels, from solving domain-specific tasks to addressing generalizable problems across disciplines. To gain a comprehensive view of visualization practice in time series, this dissertation delves into two general data types: qualitative and quantitative. For each data type, we examine how visualization helps solve domain problems and the implication of such solutions in the broader context. By combining different visualizations into one coordinated view, multiple aspects of data can be shown in conjunction with one another. To this end, we present three case studies of qualitative and quantitative visualization dashboards and close with an overarching event detection framework that can be generalized beyond the specific scenarios presented in the case studies. This dissertation contributes novel visualization designs and techniques that foster interactive systems for exploratory analysis of time-series data. From the qualitative time-series visualization perspective, the field witnessed an increasing need for a method that could effectively demonstrate the progression of topics without losing the relevant context. As a result, WordStream was created to help visualize topic evolution while providing a visually appealing representation of qualitative text data. WordStream emphasizes frequent topics that emerge from the text source, especially when they fluctuate significantly. As a natural extension of WordStream to make it more applicable to a broader demographic, we developed WordStream Maker, an end-to-end platform that automatically generates WordStream visualizations with input from raw text data. This development makes it possible for users without prior programming knowledge to create WordStream with ease. We then employed WordStream in a multiplicity of contexts to explore its potential, constraints, and adaptability in different situations, which were grounded in two case studies: Earthquake Situational Analytics (EQSA) from social media and Journal Data Dashboard from educational assessments. The findings demonstrated that WordStream was intuitive, clear, and easy-to-use to explore text entries, especially words of interest. The potential of this tool can be extended for larger real-world scenarios such as text analysis of longitudinal studies. To explore the quantitative time-series visualization realm, we proceeded with a typical case of numerical log data. Recognizing the need to leverage visual representations in malware analysis, we developed MalView, an interactive visualization platform for comprehending malware behavior. This tool provided a comprehensive visualization dashboard to represent the behavioral properties of malware classes, aiming to capture the common visual signatures of these malicious applications. Several case studies showed that MalView offered additional features and information compared to several existing visualization tools to facilitate the malware analysis process, including temporal dependencies and interactions among processes. The platform was designed to offer scalability to multiple malware families and provide a useful asset for malware experts. From the lessons learned from building qualitative and quantitative visualization dashboards, we formulated an Automated Event Detection Framework that supports identifying temporal events automatically. While events can be discerned through visual inspection, the escalating complexity of data introduces challenges that necessitate automated analysis techniques. The framework utilized the underlying graph structure from time-series data to examine the relationships among entities and extract the associated event features. Based on these identified characteristics, temporal events were detected. To evaluate the effectiveness of this method, we returned to the previously presented qualitative and quantitative dashboards to apply the framework and test its usefulness and practicality. The findings demonstrated that automatic detection mechanisms supported identifying complex events beyond immediate observation, offering a potential solution to alleviate the cognitive load associated with manual visual inspection.