Community Detection with Heterogeneous Information Networks

Many real-world systems, when represented as graphs, exhibit locally inhomogeneous edge distributions and form communities—groups of nodes with high intra-group edge density and relatively low inter-group edge density. Community detection has numerous applications, partic- ularly in social and recommendation data mining, such as detecting social events and providing recommendations. However, the noisy, dynamic, large-scale, and heterogeneous nature of real- world data presents significant challenges for community detection. In this thesis, I address four key challenges: dynamicity, scalability, multi-relational data, and heterogeneous data sources. First, I present a paper that leverages graph neural networks for detecting communities in dynamic data, addressing the issue of dynamicity. Second, I introduce a paper that hierarchically minimizes structural entropy (SE) to detect communities from large graphs in an unsupervised manner, tackling the scalability challenge. Third, I cover a paper that extends SE, originally designed for homogeneous graphs, to detect communities in multi-relational graphs. Lastly, I discuss a paper that fuses heterogeneous data sources for embedding learning, aiding in community detection-based recommendation. I conclude by summarizing the contributions made, the techniques used, and proposing future directions. The proposed methods are evaluated on publicly available Twitter, Yelp, as well as Amazon review datasets. The experiments do not involve human subjects, animals, or recombinant DNA, therefore, approvals from the Institutional Review Board, the Animal Care Committee, or the Institutional Biosafety Committee are not needed.