Unlocking the power of graphs: innovative approaches for graph mining and learning

<p dir="ltr">Currently, graph-structured data with nodes and edge links between nodes is widely used to model complex structures in real scenarios. For example, users are as nodes, and relationships between users are as edges in social networks. For brain networks, brain regions are defined as nodes and functional connectivities between brain regions are as edges. To effectively process graph-structured data for graph mining and learning, graph representation learning methods have been developed and aim to learn lowdimension vector representations for single nodes or the entire graph, i.e., node-level and graph-level representations. These methods have evolved from traditional graph embedding methods to graph neural networks (GNNs). Besides, they have widely applied in real applications, including graph anomaly detection, drug discovery, and disorder diagnosis. Although existing graph representation learning methods and their applications achieve great progress, they ignore that each graph has a different size and structure and needs adaptive graph learning ability for graph-level representations. In addition, how to learn effective graph representations to improve node-level and graphlevel anomaly detection and brain disorder analysis also needs to be considered. In this dissertation, we aim to propose a novel graph-level representation learning method and enhance graph mining in graph anomaly detection and brain disorder analysis applications. </p><p dir="ltr">First of all, we summarize different learning paradigms of existing graph-level representation methods and observe that they easily destroy the global structure of the graph and cause node information loss for graph-level representations. In addition, they cannot adaptively learn discriminative graph-level representations for different graphs because of the different sizes and structures of each graph. To solve the problems above, we design a reinforced pooling graph neural network to adaptively perform the graph coarsening process for each graph and generate the most representative coarsened graph to avoid global structure damage and information loss. Then, graph-level contrastive learning is introduced to enhance the global information preserving of graph-level representations. Secondly, I analyze that existing node-level anomaly detection methods directly use GNNs to learn graph representations for abnormal node evaluation, but over-smooth node representations led by the convolution operation of GNNs make it difficult to distinguish the difference between normal and abnormal nodes. Thus, we design a tailored deep graph convolutional network to learn anomaly-aware node representations while capturing the community structure of graphs. This aims to consider various node anomalies and alleviate the over-smooth node representations to highlight the feature representations of abnormal nodes. Thirdly, being different from node-level anomaly detection which recognizes abnormal nodes within a graph, graphlevel anomaly detection aims to recognize abnormal graphs in a graph set. It mainly faces the following challenges: 1) how to learn comprehensive graph representations to distinguish normal and abnormal graphs; 2) abnormal graphs and normal graphs have obvious number imbalance, leading to mainly learning normal graph patterns and suboptimal detection performance. To answer these challenges, we propose an endto- end graph-level anomaly detection framework by combining graph neural networks and contrastive learning. We aim to design a graph contrastive learning paradigm to enhance node-level and graph-level representations for capturing local and global graph anomalies. Finally, to apply graph mining for brain disorder analysis, we first model neuroimaging data into brain graphs and introduce knowledge distillation guided brain graph learning framework to extract important brain subgraphs, so these subgraphs can be used to enhance brain graph-level representations for disorder detection and interpret pathogenic mechanisms. Extensive experiments on real datasets verify the effectiveness of our methods, which also highlights the importance of innovative graph representation learning methods and graph mining in real applications.</p>