The Faculty Outstanding MPhil Thesis Award recognizes and rewards an MPhil student each year, whose thesis has been identified by the Faculty’s Selection Panel. The prize will be presented at the Faculty Annual Awards Presentation Ceremony to be held in March 2022.
Mr. Hou’s thesis titled “Measuring and Improving the Use of Graph Information in Graph Neural Networks” explores how to use the relation information among entities in graph data. Graphs are powerful data structures to model relations, and machine learning on graphs has demonstrated significant impact on a series of applications such as classification, prediction and recommendation. However, existing work has largely ignored the side information of graphs such as properties associated with nodes and edges. Besides, currently there is limited understanding on how much relation information that can be extracted and utilized by machine learning models.
In this thesis, Mr. Hou tackled the two problems mentioned above. First, he proposed a general framework that incorporates side information in graph data. His work makes it possible for most kinds of side information to be captured in the model to be trained and allows most existing models to be easily deployed in the framework. Second, he also introduced an analysis framework and proposed two metrics to measure the quantity and quality of the obtained relation information, which helps researchers and practitioners understand how much and how good relation information is extracted and utilized by a given model. Based on these two metrics, he also proposed a new model that can more effectively capture relation information in graph data.