Deep Community Detection for Data Analysis and Visualization

Abstract

In this thesis, we address the fundamental problem of community detection in various disciplines, including computer, social, physical, and engineering sciences. The main objective of community detection is to identify the hidden relationships between nodes in a network by identifying the communities in which the nodes are densely connected to each other. Graph mining and graph learning techniques have demonstrated remarkable performance in solving community detection problems. Recently, the Transformer has emerged as one of the breakthrough deep learning methods for natural language processing and AI applications. In this research, our main aim is to incorporate the Transformer within a Graph Convolution Neural Network (GCN) to learn encoded features and pull out the structure of the nodes, which will then be used for node classification and community detection.
We initially worked with the Cora Dataset, which includes scientific publications and citations, and achieved state-of-the-art accuracy performance for node classification. However, we encountered issues with the dataset’s representation of nodes, leading us to shift our focus to a new dataset consisting of textures found in histological images of human colorectal cancer. This new dataset includes 5,000 images of eight non-overlapping tissue types, allowing us to extract features from the images themselves rather than relying on a pre-existing feature representation.
We believe that incorporating the Transformer within a GCN will enhance the precision of node classification and community detection, achieving higher accuracy rates for these tasks. Our approach aligns with our proposed method, and we aim to achieve our primary objective of enhancing node classification and community detection through deep learning techniques.

Date of Award	Apr 2023
Original language	American English
Supervisor	Sajid Javed (Supervisor)