Files
Abstract
Node classification on networks, i.e., graphs, is an essential topic in the graph machine learning field, with ubiquitous applications in natural science, social science, industries, etc. In this dissertation, we study a few graph machine learning cases where target graph data are potentially out-of-distribution (OOD) compared to training data. Among various OOD problems, one important field is domain adaptation, where the source and target data (i.e., domains) share the same feature space but different distributions. Current domain adaptation methods for node classification only focus on the closed-set setting, where source and target domains share the same label space. Moreover, current OOD node classification methods only focus on the node classification task, leaving out the task of OOD detection (also known as anomaly detection). We first address the open-set graph domain adaptation learning problem, which can leverage source domains with rich labels to deal with classification tasks in an unlabeled target domain. Specifically, we develop an algorithm for efficient knowledge transfer from a labeled source graph to an unlabeled target graph under a separate domain alignment strategy, to learn discriminative feature representations for the target graph. Our goal is to not only correctly classify target nodes into the known classes, but also classify unseen types of nodes into an unknown class. Experimental results on real-world datasets show that our method outperforms existing methods on graph domain adaptation. Secondly, we study an OOD detection problem on graphs and classify the OOD nodes at the same time. Methods of Deep Support Vector Data Description and graph data augmentation such as Mixup are used for the OOD detection and node classification tasks, and we extensively compare our method with others on the graph OOD benchmark datasets.