Multidimensional data are commonly used to represent both structured and unstructured information. Understanding the innate relations among different dimensions and data items is one of the most important tasks for multidimensional data analysis. However, relational data patterns such as correlations, co-occurrences, and many semantic relations such as causality, topics and clusters are usually difficult for users to detect as the data are usually heterogeneous in nature, huge in amount, and contain various statistical features. Although many fundamental data analysis techniques such as clustering and correlation analysis have been widely used in various application domains, it is still difficult for users to understand, interpret, compare, and evaluate analysis results given the lack o...[ Read more ]

Multidimensional data are commonly used to represent both structured and unstructured information. Understanding the innate relations among different dimensions and data items is one of the most important tasks for multidimensional data analysis. However, relational data patterns such as correlations, co-occurrences, and many semantic relations such as causality, topics and clusters are usually difficult for users to detect as the data are usually heterogeneous in nature, huge in amount, and contain various statistical features. Although many fundamental data analysis techniques such as clustering and correlation analysis have been widely used in various application domains, it is still difficult for users to understand, interpret, compare, and evaluate analysis results given the lack of context information. Information visualization can be of great value for multidimensional data analysis as it can represent the data in intuitive ways with rich context over multiple dimensions and also support explorative visual analysis that keeps humans in the loop.

In this thesis, we introduce advanced visual analysis techniques for uncovering relational patterns in complicated multidimensional datasets including the structured multivariate data, unstructured text documents, and heterogeneous datasets like social media data that contain both structured and unstructured information. Multiple visualizations are designed for these three data types to represent relational patterns within the same or across different information facets. First, for multivariate data, we introduce DICON which is an icon-based cluster visualization that embeds statistical information into a multi-attribute display to facilitate cluster interpretation, evaluation, and comparison. Then, for unstructured documents, we design a set of visual analysis systems, Contex-Tour, FacetAtlas, and Solarmap, for topic analysis based on our proposed multifaceted entity relational data model. These systems respectively represent the multifaceted topic patterns among name entities, the multi-relational patterns within topics inside the same information facet, and the semantic relational patterns within topics across different information facets. Finally, for heterogeneous data such as twitter datasets, we introduce Whisper for visualizing dynamic relationships between users in context of the information diffusion processes of a given event. These relations contain information from three key aspects: temporal trend, social-spatial extent, and community response of a topic of interest.

To the best of our knowledge, the above techniques are cutting-edge studies of visually analyzing relational patterns in structured, unstructured, and heterogeneous multidimensional datasets. To show the power and usefulness of our study, all the proposed visual analysis systems and corresponding techniques have been deployed to real datasets and have been formally evaluated by domain experts or common users.