Graph neural network and its applications

Abstract

There has been a growing number of non-Euclidean data generated with complex interactions among the objects from different fields, including computer vision, biochemistry, and material science, which is difficult for traditional machine learning algorithms to process. Hence Graph Neural Network (GNN) has gained popularity recently since it can easily handle the data of such graph structure. GNN uses message passing of information extracted by neural network among the nodes to update the node and graph information, thus getting a better understanding by incorporating both the topology and feature space and performs outstandingly on the task such as node or graph classification and link prediction. However, there are challenges remaining for methodologies and application of GNN: firstly, it is difficult and expensive to get high-quality annotation labels for each node in node classification by GNN, but the pseudo-label of nodes generated in graph contrastive learning is heuristic and error-prone; secondly, although there have been some studies on using GNN for an organic compound such as protein, studies are lacking on how to specifically apply GNN for inorganic physics material especially considering the unique interaction in its crystalline structure. In my research, I study both challenges and propose corresponding solutions. In this dissertation, I begin by briefly describing the methodology and application of GNN. In the second chapter, I propose a dynamically denoised contrastive loss on the graph to rectify the error-prone guidance of the pseudo-label generated. In the third chapter, I use GNN on the problem of property prediction of physics materials, which is a hard problem for traditional machine learning algorithms but appropriate for GNN since orbitals in the materials have strong interactions among them. There have been some applications of GNN in Multiple Object Tracking (MOT) and Single Object Tracking (SOT). However, existing MOT algorithms, whether they use GNN or not, often request prior knowledge of the tracking targets (e.g., pedestrians) and do not generalize well to unseen categories. Thus in the fourth chapter, I propose the benchmark and protocol of Generic Multiple Object Tracking, which requires little prior information.Similarly, the current SOT algorithm is limited by small or low-quality annotated benchmarks. Hence in the fifth chapter, I propose a densely-annotated high-quality Large-scale Single Object Tracking benchmark (LaSOT) to address such issues. On the other hand, it is a great challenge for humans, even medical experts, to identify the exact type of dental implant from a radiograph image. But such pixel-level differences can be captured by Convolutional Neural Network, and high accuracy is achieved. Finally, I conclude with a discussion of future work, including the use of graph contrastive learning on physics material property prediction and Generic Multiple Object Tracking.