Modeling Structure and Behavior using Graph Machine Learning (80% Seminar)

Graph structures are ubiquitous and help us to understand many complex real-world applications. Graph Machine Learning (GML) is a branch of computer science that aims to understand a graph's hidden relations and structure and use them to provide powerful graph models for prediction, clustering, and classification tasks. Graphs are also capable of explaining complex behavior by analyzing graph models at the macro level or by studying the interaction of nodes and edges at the micro level. The theme of this seminar is to learn and understand the global and local behavioral properties of complex systems by modeling their graph structure. To achieve this, we investigate the graph structures of complex systems in two real-world domains - (i) Software Testing in Software Engineering and (ii) Cell Graph Modeling in Digital Pathology.

For tasks in the software testing domain, we use control flow graphs (CFGs) for understanding global behavioral properties such as (i) reverse engineering software requirements (metamorphic relations) and (ii) estimating code similarity. In the area of digital pathology, we generate cell graphs from digitized medical tissue images, where each node represents a cell nucleus and edges represent the spatial relationship between two cell nuclei. Cell-graphs can be used to understand tissue structure and function in histology.

In this seminar, we show that both global and local behavior properties of complex systems can be modeled from their graph structure using GML techniques. However, to achieve higher performance, it is critical to consider relevant: attributes, embeddings, topologies, and architectures.