Visual Analytics of Evolving Graphs

Abstract

Making sense of large datasets is a complex task that requires considerable effort, expertise, and time. However, the challenges are even more significant when dealing with evolving network datasets due to their dynamic, complex, and constantly changing nature. Analyzing and visualizing these datasets requires advanced techniques and tools that can handle their complexity and facilitate the identification of patterns, trends, and insights. This thesis proposes three interactive visualization systems that assist in interpreting complex network data and generating insight. Each work pursues simple visual metaphors that focus on relevant data subsets. Then, interactive systems built around these metaphors enable users to compare different temporal snapshots and capture changes in dynamic graphs.

The graph Venn diagram is a novel graph layout algorithm that analyzes the evolving content of social media streams. We first construct a knowledge graph from the social media stream represented by important people, places, and organizations. The graph Venn diagram arranges the temporal sub-graphs from the knowledge graph to emphasize the changes. We implement the KGDiff system to interactively visualize graph differences highlighting essential people or places in the last month versus the current month and those in the intersection. The KGDiff system is useful for illustrating topical shifts, identifying changes in user behavior, or discovering differences in graphs from distinct sources or geography.

Next, we propose AttentionFlow, an interactive system that visualizes networks of time series. It visualizes how attention trends of different items influence each other from large-scale graph data from YouTube recommendations (10K nodes) or Wikipedia hyperlinks (100K nodes). AttentionFlow incorporates both static and dynamic approaches to demonstrate the correlation between changes in time series and the evolving network. It aligns an ego network with the time axis of a line chart to highlight the connections among different time series within the context of the ego network. When the ego node is moved along the time axis, AttentionFlow adapts the network display to show a time-sensitive version. The static timeline allows users to compare the time series of the ego node and its neighbors and also serves as an interactive input to specify the time window. Case studies explore the impact of adding new videos to the YouTube recommendation network and visualize the effects of external/internal events on Wikipedia traffic for cultural entities.

Finally, we present the Influence Map system for visualizing academic influence via a citation network using a flower-shaped metaphor called the Influence Flower. The Influence Flower metaphor beautifully represents the intellectual influence among academic entities such as individuals, organizations, venues, and topics. The Influence Map system allows searching and curating of the Influence Flower, and also provides a visual comparison method that highlights the change in influence patterns over time. This system covers publications in all scientific domains and is open to the general public, making it accessible to a broader audience. It is also deployed plugin at arXiv.org, attracting thousands of daily queries at the time of writing. We use academic data as case studies to visualize researchers' career trajectories over time, the interdisciplinary profile of an institution, and the evolving topical trends in conferences. Show more