Visual Analysis Of Association Data Based On Graph Embeddings

As an important research object of data analysis,association data exists in various fields of life.It is usually represented as graph,node represents entity,link represents inter entity Association.Effective analysis and exploration of graphs can help people to discover knowledge hidden in complex information and assist decision-making.Due to the complexity of large-scale graph data structure in practical application,it is difficult for users to directly and efficiently analyze graph structure,so they need to use artificial intelligence technology such as graph embeddings.Similarly,graph embedding models also have problems of poor interpretability and different model performance,which makes it difficult for users to compare and select.It is difficult to accomplish the task of graph analysis by human intelligence or machine intelligence alone.Therefore,based on the National Natural Science Foundation of China and National Key R&D program of China,this study combines graph embeddings with visual analysis technology,helps users to compare graph embedding models by visual analysis,selects appropriate models and optimizes models according to the analysis requirements,and helps users to explore the implied relationships in data efficiently.The following achievements have been achieved:First,a normalized form of graph embedding models structure is proposed.The graph embedding models are divided into three parts: graph sampling method,neural network structure and loss function.The neural network is further strengthened to supplement the decoder for the neural network with only encoder,so as to make more loss functions act on the neural network and enhance the ability of the model to retain multiple graph structure features.Using tensorflow framework,according to the standardized structure,the deepwalk,nod2 vec,struc2vec and SDNE models are reconstructed and realized.It is convenient for users to choose different graph sampling methods,neural network structure and loss function to construct new graph embedding models in combination form,so as to achieve better graph embeddings effect.Secondly,a visual analysis method and system GEMVis is proposed to evaluate,compare and improve the graph embedding models.In this method,nine node metrics such as degree,eccentricity and closeness are used to describe graph structure features.Interactive methods and color mapping are used to help users explore the relationship between embedding vectors and node metrics.Neural network regression model is used to measure the ability of graph embedding models to retain graph structure features.The method of multi view collaboration is used to support the qualitative and quantitative comparison and analysis of multiple embedding models from three aspects: embedding vector,node measurement and original graph structure.It supports users to adjust model parameters and construct a new graph embedding model in the form of combination to meet users' pursuit of metrics at specific nodes.Finally,three cases are given to illustrate the effectiveness of this method.Thirdly,a visual exploration method of food safety risk factors based on graph embeddings is proposed.Taking the detection data of import and export food hazards as the research object,taking food as the node,the link is established according to the similarity of detection of hazards in food,and the association chart is constructed.Graph embedding method is used to transform graph into a set of vectors,and K-means clustering method is used to gather nodes into several subsets.In this paper,a new overview map is designed,which uses data portrait technology to depict the surrounding structural features of each sub centralized node and show them to users.Users can select the set of interested nodes to generate other auxiliary views and further analyze the external attribute characteristics of nodes.From the perspective of graph,help researchers to explore the hidden risk factors in the detection data of imported and exported food hazards,and find out the food and hazards that need attention.