A comprehensive study on spectral clustering

Abstract

Spectral clustering is currently a widely used method for community detection. This Final Year Project (FYP) researched and learned on spectral clustering comprehensively in three main perspectives. Firstly, a guideline on the selection of similarity matrices, adjacency matrix and Laplacian matrix, under different conditions is proposed through a large number of simulations in Chapter 2. Next, an improved spectral clustering method with more general input metrics is investigated with satisfying performance in Chapter 3. Lastly, two methods for the number of clusters K estimation are introduced and compared in Chapter 4. The better method is suggested based on simulation results under different number of nodes, clusters and blockmodels. Overall, the project overcomes the limitations of conventional spectral clustering algorithms towards similarity metrics and distance metrics, and constructs a complete flow to carry out spectral clustering.