•   When: Tuesday, June 22, 2021 from 11:00 AM to 01:00 PM
•   Speakers: Matthew Revelle
•   Location: Virtual
•   Export to iCal

Social networks are defined by relationships between people and permeate all aspects of human life. Improving our understanding of the structure and dynamics of networks enhances our knowledge of many human systems. In my dissertation, I present novel techniques and methodologies for the tasks of community detection, role discovery, and community evolution prediction as well as an analysis on temporal artifacts that may occur when constructing social network datasets. For this presentation, I will focus on my contributions to community detection and community evolution prediction.

Community detection in networks is a broad problem with many proposed solutions. Existing methods frequently make use of edge density and node attributes; however, the methods ultimately have different definitions of community and build strong assumptions about community features into their models. I propose a new method for community detection, which estimates both per-community feature distributions (topics) and per-node community membership. Communities are modeled as connected subgraphs with nodes sharing similar attributes. Nodes may join multiple communities and share common attributes with each. Communities have an associated probability distribution over attributes and node attributes are modeled as draws from a mixture distribution. The method includes two basic assumptions about community structure: communities are densely connected and have a small network diameter. These assumptions inform the estimation of community topics and membership assignments without being too prescriptive.

Communities in social networks evolve over time as nodes enter and leave the network and their activity behaviors shift. I present a novel graph attention network model for predicting community evolution events based on group-node attention. Group-node attention enables support for variable-sized inputs and learned representation of groups based on member and neighbor node features, including temporal information. Existing work on community evolution prediction has focused on the development of frameworks for defining events while using traditional classification methods to perform the actual prediction. It is my hope this work on a prediction model for community evolution events will prompt the development of additional novel prediction models.