Graph Neural Network Based On Local Subgraphs And Interaction Modeling

Graph neural networks are promising architectures for learning and inferring graphstructured data:as small as cells and molecules,as large as social networks,and even the behavior of celestial bodies.Following the great success of deep neural network end-to-end solutions in processing structured data(image,speech,language),the analysis,understanding,modeling and prediction of highly unstructured,non-Euclidean spatial relational graph data have been received great attention and has shown important theoretical and applied value in the fields of information technology,natural science and engineering.At present,GNN(Graph Neural Networks)is the most advanced neural architecture for graph mining tasks in the field of artificial intelligence.However,in the face of graphs with ever-changing topology,number of nodes,and feature distribution,graph neural networks have certain defects in dealing with complex topologies:(1)How to use deep neural network technology for modeling and computing of graph-structured data,generating efficient and compact graph representation with fixed dimension and reliable reflection of topology information.Graph neural networks have limitations in characterizing the constituent units(local subgraphs)of a system.Conventional graph convolution operators use nodes as processing units.Although they can describe the "central subgraph" and distinguish special graph structures,they cannot cover general subgraphs.Moreover,the weighting of node features cannot clearly describe the topological relationship,let alone describe the relationship between individuals;(2)how to build a fitting model to accurately predict the target properties of the graph.Local subgraphs are discrete expressions,so the existing graph mining usually adopts the greedy enumeration characterization method,which can only deal with subgraphs of extremely small size.All subgraphs of smaller size greatly limit the coverage of subgraph changes and lead to overfitting;(3)how to quantitatively correlate the structural features of the attribute graph with the functional attributes in an interpretable way,so as to provide experts concise and novel rules and scientific hypotheses to facilitate rule mining and knowledge discovery.In contrast,complex system research seeks commonalities from multiple types of system instances and abstracts them into universal rules.The scientific ideas accumulated by this kind of cross-disciplinary and cross-domain inductive verification provide a rare reference for the structural design of graph neural networks,which should be further explored.The interdisciplinary and interdisciplinary scientific ideas of complex systems provide important inspiration for the design of graph neural network models.However,this association has not been fully exploited,so there is still room for deep cross-integration of complex systems science and graph neural networks.This paper takes graph neural network algorithm theory as the research object,and overcomes its lack of effectiveness in characterizing complex topological structures by introducing the characteristics of data local mining(local subgraph individual representation)and global modeling(interaction modeling)in complex system research,the lack of stability of network training generalization,and the lack of interpretability of the model,so as to improve the design of graph neural network architecture.In summary,for solving the problem of local subgraph representation and its interaction in graph neural network research,the main work and innovations of paper are as follows:1.Aiming at the problem of node classification,a node classification method based on adaptive structure fingerprint is proposed.Despite numerous successes,how to make good use of the structural information in the GAT remains a challenge.One is that the weighting coefficients in GAT only depend on the structural information of the graph,which is different from the actual situation,in which the relationship between nodes may also be derived from their characteristics.Second,GAT mainly uses node features and first-order neighbors to compute attention.However,using higher-order neighbors negatively affects its performance,which is closely related to the over-smoothing nature of GNNs.This chapter associates each node with a "structural fingerprint" composed of its higher-order neighbors.Each fingerprint is a subgraph with adaptive,nonnegative weights,reflecting rich local structural details,enabling message passing among higher-order neighbors while eliminating unwanted inter-class interactions,verified from a node embedding perspective The advantages of the local subgraph as the basic research unit in the field of graph mining are presented.2.Aiming at the problem of graph classification,a graph classification method based on subgraph structure flags and topology-preserving graph pooling is proposed.Although the graph neural network has achieved vigorous development in recent years,its core pooling mechanism still follows the original "node convolution-compression pooling" framework,which can deal with the topology of complex attribute graphs,training stability and representation ability,And there are some deficiencies in the interpretability of the model.The existing pooling operators are prone to loss of topological information and affect the ability of topological representation.The hyper parameterization of neural networks can enhance the capacity and fitting ability of the model,but it can also lead to overfitting and model instability.A simple neural network architecture"Subgraph Structure Flags and Topology Preserving Graph Pooling" or SLIM is proposed here.The key idea is to compute a set of structure flags and use local subgraph instances as the basic processing unit(rather than node neighborhoods).Feature weighted average,thus providing a richer context for the topological identity description of nodes,improving the pooling layer.3.Aiming at the drug-drug interaction problem,a drug-drug interaction prediction method based on substructure signature learning is proposed.A drug is just an entity composed of different chemical substructures(functional groups).In the existing methods for predicting drug interactions using substructures,each node is considered as the center of the substructure,and nodes adjacent to each other end up can become the center of similar substructures,leading to redundancy.At the same time,the huge differences in structure and properties between compounds can also lead to irrelevant pairings,resulting in the inability to integrate a lot of information,and this heterogeneity will have a negative impact on the prediction results.In order to solve these problems,a new method for DDI identification based on substructure,SIM-DDI,is proposed here.By extracting useful information from the local subgraph around the drug,the method can effectively use the substructure to assist in predicting drug side effects,while simultaneously Similar substructures can be pooled together using deep clustering algorithms,so that any individual subgraph can be reconstructed by this global set of signatures.In addition,a Co-attention mechanism was developed to model drug-drug interactions,generating signal intensity scores for each class of drugs to circumvent the noise caused by heterogeneity.SIM-DDI is evaluated on a generic dataset and improves the performance of DDI prediction compared to state-of-the-art methods.