Graph Neural Network Based Heterogeneous Graph Representation And Application For Trust Evaluation

Heterogeneous information networks(heterogeneous graphs)have multiple types of nodes and links,and can model the rich structural and semantic information of complex systems in a more comprehensively way.However,there are always some troubles in complex systems,such as missing description or having no access rights for some types of entities.These realistic factors undermine the integrity of the modeling process,making some types of nodes have no attributes in heterogeneous graphs.The existing attribute completion methods are only for homogeneous graphs with a small number of missing attributes,and are not suitable for heterogeneous graphs with complex network structure and multiple types of nodes having completely attribute missing.And the existing heterogeneous graph methods usually perform simple preprocessing for missing attributes(such as mean value imputation and constant vector imputation).The imputation-based methods fail to consider the properties of heterogeneous graphs and the complexity of missing attributes.These preprocessing methods bring large errors to the attributes,which will seriously affect the performance of models.To this end,we focus on designing a heterogeneous graph analysis method for complex scenarios with missing attributes,and further consider the characteristics of real scenarios,to propose a new heterogeneous graph neural network method which can be applied to the trust evaluation task.First of all,we propose a new heterogeneous graph neural network(HGNN)method that integrates the two processes of attribute completion and network representation learning together in a unified framework.In order to make full use of the valuable information in the network,we design a topology-guided attribute completion strategy.This strategy uses the rich semantics in the heterogeneous network to capture the highorder semantic relationship between nodes and attributes,and realizes learnable attribute completion.We further integrate it into the process of network representation learning to achieve a task-guided attribute completion,that is,to design a heterogeneous contrastive learning strategy for the overall framework.In this way,attribute completion and network representation learning can be optimized together and promote each other.In addition,we also design a new HGNN encoder,which uses a grouping aggregation mechanism to encode the information in a meta-path,and intelligently aggregates the semantic information between multiple meta-paths.Comparative experiments with eight advanced algorithms show that our proposed method can effectively complete missing attributes and learn node representations of heterogeneous graphs,and has superior performance.Secondly,trust evaluation aims to predict the trust level between two users,which is of great significance in user collaboration and secure communication.In this paper,we analyze the characteristics of application scenarios and design a HGNN-based method for trust evaluation.We fully consider the characteristics of missing attributes in social networks and modify the attribute completion process.We also analyze the transferable and asymmetrical characteristics of social trust,and extend the idea of knowledge graph embedding to simulate the asymmetric transfer process of social trust.Finally,we design a learnable aggregation mechanism to learn complex trust interactions,thus capturing composability of social trust.The prediction and evaluation of social trust are based on users’ embedding.We compare the proposed method with four types of representative trust evaluation methods.The experimental results show that our proposed method is more efficient in solving trust evaluation problems.The ablation analysis experiment further verified the validity of the key design in our method.