Extracting Insights from Differences: Analyzing Node-aligned Social Graphs

Social media and network research often focus on the agreement between different entities to infer connections, recommend actions and subscriptions and even improve algorithms via ensemble methods. However, studying differences instead of similarities can yield useful insights in all these cases. We can infer and understand inter-community interactions (including ideological and user-based community conflicts, hierarchical community relations) and improve community detection algorithms via insights gained from differences among entities such as communities, users and algorithms. When the entities are communities or user groups, we often study the difference via node-aligned networks, which are networks with the same set of nodes but different sets of edges. The edges define implicit connections which we can infer via similarities or differences between two nodes. We perform a set of studies to identify and understand differences among user groups using Reddit, where the subreddit structure provides us with pre-defined user groups. Studying the difference between author overlap and textual similarity among different subreddits, we find misaligned edges and networks which expose subreddits at ideological 'war', community fragmentation, asymmetry of interactions involving subreddits based on marginalized social groups and more. Differences in perceived user behavior across different subreddits allow us to identify subreddit conflicts and features which can implicate communal misbehavior. We show that these features can be used to identify some subreddits banned by Reddit. Applying the idea of differences in community detection algorithms helps us identify problematic community assignments where we can ask for human help in categorizing a node in a specific community. It also gives us an idea of the overall performance of a particular community detection algorithm on a particular network input. In general, these improve ensemble community detection techniques. We demonstrate this via CommunityDiff (a community detection and visualization tool), which compares and contrasts different algorithms and incorporates user knowledge in community detection output. We believe the idea of gaining insights from differences can be applied to several other problems and help us understand and improve social media interactions and research.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/149801/1/srayand_1.pd