Information visualization has emerged as a very active research field for multivariate and relational data analysis in recent years. It turns complex and abstract data such as demographic data, financial data, social networks, and paper citations into visual representations, and then users can exploit interactive computer graphics techniques and human visual capabilities to gain insight into the data. Parallel coordinates and graphs are two well-established methods in information visualization. However, when data become very large, the effectiveness of both methods is dramatically reduced as tens of thousands of lines can easily overwhelm the display and the resulting visual clutter will obscure any underlying patterns. Thus, clutter reduction for parallel coordinates and graphs is a ve...[ Read more ]

Information visualization has emerged as a very active research field for multivariate and relational data analysis in recent years. It turns complex and abstract data such as demographic data, financial data, social networks, and paper citations into visual representations, and then users can exploit interactive computer graphics techniques and human visual capabilities to gain insight into the data. Parallel coordinates and graphs are two well-established methods in information visualization. However, when data become very large, the effectiveness of both methods is dramatically reduced as tens of thousands of lines can easily overwhelm the display and the resulting visual clutter will obscure any underlying patterns. Thus, clutter reduction for parallel coordinates and graphs is a very important research problem in information visualization.

In this thesis, we introduce visual clustering as a new approach for clutter reduction and pattern detection. Compared with traditional clutter reduction methods such as filtering and brushing, visual clustering can enhance and reveal interesting patterns in the data while preserving the context. For parallel coordinates, we present a force-based optimization method to bundle polylines by adjusting their shapes, and a splatting framework to reveal features with animations. For graphs, a geometry-based edge grouping approach and an energy-based hierarchical visual clustering scheme are proposed. The effectiveness of these methods has been demonstrated through extensive experiments using both synthetic data and datasets from real applications.