Research On Deep Learning Based Graph Embedding Models And Applications

Graphs are common information carriers in complex systems.It represents complex relationships,such as social networks and criminal networks.Graphs contain many valuable information,so it is of great significance to analyze and mine graph in an efficient way.The graph embedding models based on deep learning have powerful representation ability.It can effectively extract the nonlinear features and generates low-dimensional vectors that retain the original graph information.The graph convolutional autoencoder is commonly used deep learning model in recent years.It uses information transfer between nodes to generate low-dimensional embeddings that preserve deep neighborhood features,and improves performance on graph analysis tasks.However,many graph convolutional autoencoders suffer from limited ability to retain original graph information and performance decline with increasing depth.We propose two models to address these problems and apply them to criminal network analysis.To address the limited ability of graph convolutional autoencoder to retain original graph information,we propose Multi-channel Graph Convolutional Autoencoders(MC-GCAE).Firstly,we design specific and consensus convolutional encoders to extract attribute features,topology features and their association.Secondly,we design symmetric convolutional decoders to recover the encoding process.Thirdly,we introduce reconstruction loss,local constraint and consensus constraint to optimize embeddings generated by different encoders.Finally,multiple embeddings that contain different information are fused to generate representations.The results show that the multi-channel approach adopted by MC-GCAE can retain richer graph information in lowdimensional embeddings,and improves performance in node classification,node clustering and visualization tasks on public datasets.To address the performance decrease of graph convolutional autoencoder with increasing depth,we propose graph embedding models OS-Se VAE and OS-Se AE using One-Shot aggregation autoencoder and second-order information.Firstly,we use One-Shot aggregation and exponential linear unit function to improve the gradient update of the deep model and avoid oversmoothing.Secondly,we introduce a regularization term in the loss function to prevent the overfitting of the parameters due to increase in depth.Finally,we propose a graph convolutional layer to extract second-order neighborhood features,which enhances the representation capability.The results show that OS-Se VAE and OS-Se AE can effectively improve the deep model performance,retain more structural information,and improve performance in link prediction task on public datasets.To address the limited ability of police data analysis based on manual analysis and statistical algorithms to handle large-scale criminal network,we use graph embedding models to mine criminal network information.Firstly,we pre-process information of criminal organization,and construct a criminal network.Secondly,we use graph embedding models to extract criminal network features,and generate low-dimensional embeddings.Finally,we use embeddings to predict identities and relationships of criminal organization members.The results show that our graph embedding models outperform baselines in accuracy and stability.Meanwhile,it can provide useful information for intelligence analysis.