Research On Evaluation And Generation Of Graph Layout Based On Graph Neural Network

In the field of graph layout generation and evaluation,most traditional algorithms suffer from disadvantages such as complex computation,high time cost,and tedious parameter tuning.With the rise and application of data mining and artificial intelligence,some scholars have introduced deep learning techniques into graph layout and evaluation,achieving good results.However,most existing deep learning-based graph layout evaluation methods use CNN(convolutional neural network)models to evaluate the quality of graph layout images,which results in low accuracy due to the complexity of graph structure,disorderliness of layout,and inconsistency of images.At the same time,existing deep learning-based graph layout generation methods suffer from low model dataset applicability,insufficient sample learning capability,and single layout style.To address these problems,this thesis explores the research work of graph layout evaluation and generation based on graph neural networks,with the following main achievements:1)This thesis proposes a graph layout evaluation framework based on graph neural networks.The framework includes objective and subjective evaluation networks that share a graph neural network-based layout feature extraction model.This model first normalizes the shape of the layout using Modified PCA(modified principal components analysis),then samples the neighborhoods of each node,and uses the subgraph features of the sampled points as the model input.Finally,it uses multi-layer graph convolution and pooling operations to obtain the global features of the input layout.Experimental results show that compared to the CNN method,the graph layout evaluation model in this thesis can achieve higher evaluation accuracy.2)This thesis proposes a graph layout generation model based on an encoderdecoder architecture.The model takes category labels and graph structural information as input and utilizes the category labels to control the style of the generated layout.In addition,a loss function is proposed in this thesis to fit the sample distribution,which can not only learn the overall layout structure of the graph but also maintain the beauty of the local layout structure.Experimental results show that the proposed graph layout generation model can be applied to graphs with different structures and can generate multiple styles of graph layout effects.3)This thesis designs and implements a graph layout visualisation system.The system provides users with a user-friendly graphical interface to display the layout effects of various traditional graph layout algorithms,assisting users in selecting suitable algorithms and parameters to generate a sample set of graph layouts.