Research On Dynamic Network Embedding Based On Deep Learning

In the real world,many systems are represented by a network,such as social networks,Internet,collaboration of scientists and power systems.However,the original network data has the characteristics of redundancy,correlation,and large-scale,which increases the requirements for algorithm efficiency and affects the effect of data processing.Network embedding learning compresses the original network data into a low-dimensional space for representation,which not only reduces the calculation cost,but also improves the effect of network data mining tasks(such as link prediction,network reconstruction,network stability,community stability,etc.).However,the previous network embedding methods mainly focused on static networks that were not suitable for dynamic networks with evolutionary properties.The existing deep learning methods for dynamic network embedding still have some problems and challenges: the low stability and high computational complexity.This article aims to solve those two issues in the framework of deep learning for of dynamic network embedding.First,this paper proposes a community-aware dynamic network embedding learning(CDNE)to improve the stability by using a deep autoencoder.CDNE uses a combination of first-order information and second-order information to better represent the structures of the original network,thus improving the effectiveness of network prediction tasks(link prediction,network reconstruction).Moreover,CDNE proposes a joint optimization model with a novel objective which maximally conserves the global structures,local structures and community evolutions.This optimization model would improve the stability of dynamic networks embeddings.Experimental results show that compared with some current state-of-the-art network embedding learning methods,the proposed method achieves better performance in prediction tasks(link prediction,network reconstruction)and stability tasks(network stability,community stability).Second,this paper proposes a sparse deep autoencoder model(called SPDNE)to decrease the computational complexity of current dynamic network embedding methods.Specifically,this paper proposes a sparse autoencoder structure evolution algorithm by using a simulated annealing algorithm to find the unsupervised trained autoencoder sparse topological structure for dynamic network embeddings.This model is validated by three classical deep network embedding methods on real-world dynamic networks.The results show that the proposed model can greatly reduce the computational complexity of dynamic network embedding methods while maintaining the effectiveness on network tasks(link prediction,network reconstruction).