Graph-based Semi-supervised and Unsupervised Local Clustering
Data Science Seminar
Zhaiming Shen (Georgia Tech)
Abstract:
Semi-supervised local clustering focuses on identifying specific substructures within a large graph without requiring a full understanding of the entire graph. In this talk, I will present a method to find the target cluster by solving a sparse solution to a linear system tied to the graph Laplacian. By applying local clustering methods iteratively, we can uncover all community structures within the graph. Extensive experiments across diverse datasets demonstrate the effectiveness of our approach. Moreover, the framework can be generalized in two key directions. First, by reducing the label ratio, or even eliminating labels entirely, we can extend the problem to an unsupervised setting. Second, we can accommodate the presence of nodes that do not belong to any underlying clusters, broadening the applicability of our method.