Abstract
<p>Graphs and network structures are powerful tools for modeling and analyzing complex systems, such as power distribution networks, water distribution networks, and communication networks. By analyzing the topological features of a graph, we can better understand how the nodes in the network are connected and how they interact with one another. This is especially important in anomaly detection, as anomalies often result in changes to the underlying structure of the network. Graph-based models provide a way to represent complex data in a structured and organized manner, allowing us to gain insights into the behavior of the system being studied. <br />
Recently, Graph Neural Networks (GNNs) have shown promise in detecting anomalies in graph-structured data. In this dissertation, I propose a framework for anomaly detection with Graph Neural networks (GNNs) that explicitly leverages the graph topology and node attributes. In addition to node attributes/features, we incorporate graph topological features, such as centrality and clustering coefficients, to improve the performance of our model. Our experiments demonstrate that adding these topological features improves the accuracy of anomaly detection in attributed networks compared to traditional autoencoders that do not simultaneously consider graph topology and node attributes. Our results suggest that incorporating graph topology (represented by structural components and node attributes) is a key factor for effective anomaly detection with GNNs.</p>