Neural Architectures for Searching Subgraph Structures

With the development of new neural network architectures for graph learning in recent years, the use of graphs to store, represent and process data become more trendy. Nowadays, graphs as a rich structured form of data representation are used in many real-world projects. In these projects, objects are often defined in terms of their connections to other things. There are many practical applications in areas such as antibacterial discovery, physics simulations, fake news detection, traffic prediction and recommendation systems. While there are a growing number of neural graph representation learning techniques that allow one to learn effective graph representations [1-3, 5], they may not necessarily be appropriate for the task of searching over graphs for several reasons: (1) nodes within a graph consist of a set of attributes, which are the subject of the search process. However, the number of unique attributes on the graph compared to the number of attributes on each node is extremely sparse; therefore, this makes it very difficult to learn effective graph representations for such sparse information; (2) graph neural networks are capable of generating rich embedding representations for nodes and entities in graph however, depending on the downstream task, the embedding vectors may perform not well as expected. Therefore, researchers need to apply solutions such as custom loss functions or pre-training tasks to adapt mentioned architectures for their specific task. Search in the graph as a downstream task follows the same trend and therefore a tailored graph neural network representation is needed to particularly address the need for a rich embedding representation for the sole purpose of subgraph search. My work focuses on searching for subgraph structures over both complete and incomplete heterogeneous graphs. The significance of my research direction lies in the fact that exact subgraph search is an NP-hard problem and as such existing methods are either accurate but impractically slow, or efficient yet suffering from low effectiveness. With a focus on learning robust neural representations for complete and incomplete graphs, my research focuses on developing search methods that are both effective and efficient. Specifically, my research addresses the following research questions: RQ1) Would it be possible to design and develop graph representation learning methods for heterogeneous graphs that can generate effective embedding vectors from a heterogeneous graph and support effective and efficient subgraph search? RQ2) Whether it would be possible to address the issue of graphs with varying degrees of missing values. Incomplete graphs suffer from missing attributes and/or missing edges. I explore the design of robust graph representation learning models capable of effectively searching in light of missing information. RQ3) Can efficient and effective derivations of my subgraph search methods be used to address practical applications in real-world domains such as team formation, and keyword search over knowledge graphs? Research conducted as a part of my Ph.D. has so far focused on identifying and retrieving subgraphs over complete heterogeneous graphs. I have used the Team Formation problem as a case study in order to evaluate my work [4]. As the next step, I am planning to focus on expanding my research to include incomplete graphs and investigate methodologies to craft optimized graph representations for the specific task of searching on graphs.