| A graph is an irregularly structured data type that consists of a set of nodes and a set of edges.Nodes typically represent a certain type of entity,while edges are used to connect two nodes,representing the interrelationships between entities.In real life,many data can be represented by graphs,such as biological molecules,protein structure,financial trading networks,etc.Graph classification and graph based node classification are two very important tasks related to graph data.Graph classification refers to the labeling of discriminative graphs.Due to the significant differences in the topological structure of each graph,the processing of graph classification tasks is challenging.Graph based node classification involves learning more robust representations of entity nodes through the use of graph structures,and then distinguishing the category to which each entity belongs.Unlike in graph classification where the graph structure is always given,in graph based node classification,there may be an initial graph structure or we may need to construct the graph ourselves.For graph classification tasks,existing methods can be divided into kernel methods and deep learning based methods.Deep learning based methods typically utilize Graph Neural Network to extract graph representations and generate classification results based on these graph features.However,these GNN based methods usually need to improve the Receptive field by superimposing multiple layers of GNN,which easily leads to(1)over smoothing and(2)the introduction of too many noise nodes.For graph based node classification,it can be divided into static graph based methods and dynamic graph based methods based on whether the structure of the graph has changed.The dynamic graph method can adaptively adjust the graph structure based on the dataset,thus better mining the internal connections of the data.In dynamic graph based methods,there are also methods based on dynamic ordinary graphs and dynamic hypergraphs according to the type of graph.Dynamic hypergraph based methods usually use Hypergraph Neural Network to extract node representations.However,(3)existing dynamic hypergraph neural networks ignore the characteristics of hyperedges and cannot adjust the number of hyperedges,thus unable to fully explore the structure of hypergraphs.For problems(1)and(2),this paper proposes a novel multi-scale Graph Classification with Shared Graph Neural Network.By proposing a multi-scale graph coarsening framework and parameter sharing mechanism,MSFG can effectively extract high-order structural features of the graph without increasing the number of GNN layers.For problem(3),this paper proposes a novel Total Dynamic Hypergraph Neural Network.By learning the feature distribution of hyperedges and sampling to obtain hyperedges and dynamically updating the hypergraph,TDHNN can adjust the number of hyperedges while optimizing the hypergraph structure.The effectiveness of MSFG and TDHNN was experimentally verified. |
