Research And Analysis Of The Work Of Heterogeneous Motifs In Graph Representation Learning

Many data in real-world scenarios should be described and studied by modeling them as graphs.Due to the variety of data types,their internal structures are becoming more and more complex.In the case of inconsistent data types,they are usually represented as heterogeneous information networks.Heterogeneous information network,also known as heterogeneous graph,differs from homogeneous graph in that heterogeneous graph data structure with multiple node types or multiple edge types for portraying different types of objects and their interactions with rich structural and semantic information,which provides a more effective modeling tool and analysis method for graph data mining.Heterogeneous information networks contain different types of entities with multiple complex interactions between them,which together provide a more detailed description of the semantic structure than in homogeneous graphs.To capture the complex relationships in heterogeneous networks,researchers have proposed mining structural patterns in the network to describe the underlying semantic structural relationships,and one class of network schemas is the graphlet also named motif.Motifs in heterogeneous graphs are called heterogeneous motifs,which are induced subgraph structural patterns with specific semantics in heterogeneous graphs.There are two advantages of using motifs,especially heterogeneous motifs in heterogeneous graphs,for representation learning:1.Heterogeneous motifs can capture complex neighborhood structures and higher-order structural relationships in the network.2.Heterogeneous motifs can perform accurate role recognition of nodes based on contextual node types and structural patterns.In recent years,more and more graph researches incline to use substructures such as tuples as a cornerstone for studying related work on graphs,e.g.,in the field of representation learning there exist structures such as meta-paths or graphlets/motifs for learning higher-order of locally complex relationships of nodes in graphs to perform network representation tasks.In the course of studying heterogeneous graph representation learning,two limitations have been identified in prior work on the usage of motifs:1.When mining motifs in heterogeneous networks,some motif structural patterns are usually predefined based on subjective experience to describe the semantic relationships in the network.There is a tendency to emphasize a certain part in defining motifs,which ignores the integrity of the mining motif patterns.2.When using heterogeneous motifs,the representation ability of different motifs is different,then exploiting the core motif and combination of them in the network and judging the effectiveness of them in graph representation learning tasks are issues worth studying.To address the above issues,the main work of this paper is as follows.1.Propose a non-predefined objective and complete algorithm for mining the motif patterns in heterogeneous networks for eliminating the subjective influence brought by relying on domain knowledge when using heterogeneous motifs for network representation tasks.A detailed analysis of the semantics structure and statistics of the mined motifs is also presented and verify the feasibility.2.A framework process for judging the effectiveness of heterogeneous motifs is proposed and based on which the performance of such motifs in participating in network representation task is verified.Experimental study and analysis of the effect of single and combined graph element expressions are conducted separately to get the core structures and efficient representation.The idea of graph fusion is used in the representation of multiple graph elements while adding weight modulation vector to investigate the representation effect of combined graph elements.In addition to this,the problems of augmentation between paired motif patterns and similarity caused by information redundancy are experimented and analyzed.Experiments based on classification and clustering of target nodes on the dataset show that integrity-mined heterogeneous motifs are better in representation compared to predefined network schemas.