Interactive and Visual Analysis of Networks

Network clustering methods play a key role in the study of large and complex systems in many fields of science. One popular approach for network analysis is community detection, which reveals sets of nodes that are densely connected within rather than between modules. However, it is at the intermediate steps towards graph clustering and community detection that visual analysis can provide essential insights into uncertainty and variability of clustering decisions. To date, there exists no principled framework that is able to combine graph clustering techniques, exploration of the parameters of these techniques, and interactive visualization. It is only when these pillars of network analysis & visualization are tightly coupled that information extraction and discovery will be optimal. The IVAN project will provide an open-source software framework that will combine research in advanced network clustering algorithms, novel dynamic network visualizations, and visual analysis for the interactive exploration of network clusterings algorithms, to allow network analysts to compute meaningful and sound clusterings in a generic way, but applicable to a large range of domains. The framework will be applied to a broad range of challenging network data with varying sizes, connection densities, and organizational complexities. In particular, the methodology will be validated on synthetic datasets with known ground truth, such as Stanford Large Network Dataset (SNAP), but also on social networks, a large coauthorship network, and EEG and fMRI brain networks. Our tools will be provided as open-source software libraries, including an interactive visualization tool for end- user applications. This project is a collaboration between researchers of INRIA France, EPFL Lausanne and the University of Vienna.

Interactive and Visual Analysis of Networks (IVAN) was an interdisciplinary project to create visual analysis systems for exploring dynamic or time-dependent networks (from small to large scale). In this project, we collaborated with domain experts from various domains, including medicine and digital humanities, who work with network data. We conducted several creativity workshops to characterize their domain problems and ideate solutions. We reported on our experience running these workshops and extended existing guidelines in a research paper. Our collaborators at EPFL developed a novel algorithm for zooming on specific network parts. This algorithm produces hundreds of scatterplots which makes analyzing and comparing them difficult. To facilitate this process, we developed a novel technique called "ClusterCover" that allows for assessing clusters of scatterplots by summarizing the cluster content with four different levels of visual abstraction (confidence ellipses, convex hulls, concave hulls, and all data points). We also collaborated with neuroscientists who work on EEG data from epilepsy patients. While the results from our requirement analysis showed that they primarily use mathematical analysis to explore the data, we found that visual exploration before statistical testing could improve this process. Hence, we developed two high-fidelity prototypes for analyzing multi-frequency EEG data. The evaluation of the prototypes with domain experts and lay people showed the great potential of our visual analysis tools.