Combining automated analysis with interactive visualizations for spatio-temporal data analysis

Abstract

The rapid advances in sensing technologies and large-scale computing infrastructures lead to explosive growth in data. Spatio-temporal (ST) data, as a ubiquitous type of data, is increasingly collected and extensively studied in various scientific domains such as geology, climatology, sociology, and transportation science. This type of data is distinct from others due to the simultaneous presence of spatial and temporal dimensions, which substantially increases analysis complexity. Purely automatic data analysis techniques are insufficient to handle such complexity immaculately. Humans not only have inherently good senses for perceiving space and time but also possess creativity, flexibility, and domain expertise. Hence, an appropriate method that involves these human traits into automatic data analysis will be tremendously helpful. In this thesis, we introduce three novel visual analysis techniques for ST data analysis to demonstrate the benefits brought by the combination of automatic data analysis techniques with interactive visualizations. First, we study how to solve a multi-criteria decision-making problem in the spatial-temporal context that involves a vast solution search space. We use optimal billboard location selection as our application scenario and propose SmartAdP. This system integrates a novel visualization-driven data mining model with tailored data index mechanisms to facilitate efficient solution formulation. Several well-designed visualizations are also put forward to support optimal solution identification. Second, we investigate how to detect and examine anomalous events hidden behind a large number of spatial time series. We use air quality analysis as our primary application scenario and present AQEyes. The system contains a unified end-to-end tunable machine learning pipeline that supports quick identification of anomalous air pollution events. A set of novel visualization techniques are presented to facilitate efficient exploration of air quality dynamics and examination of detected anomalous events. Third, we research how to quickly identify ST patterns hidden within the subsets of large-scale multidimensional ST datasets. We propose a novel tensor-based algorithm to allow automatic slicing of data into homogeneous partitions and extracting latent patterns in each partition for comparison and visual summarization. Based on the algorithm, we further develop TPFlow, a system supporting a top-down, human-steerable, and progressive partitioning workflow for level-of-detail multidimensional ST data exploration. The effectiveness and usefulness of the above techniques are validated through case studies on real-world datasets and interviews with domain experts. The proposed techniques are not limited to the presented example application scenarios. They can be easily adapted to other applications with similar problems as well.

Date of Award	2019
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology