Application And Research On Dynamic Evolving Network Based On Representation Learning

In recent years,with the development of artificial intelligence,deep learning as an important technology has been used in frontier fields such as natural language processing and computer vision.At the same time,dynamic evolving networks widely exist in our real life,including social networks,scientific research citation networks,and transportation networks.The study of dynamic networks has also become a hot research topic in academia.At present,there are a large number of researches on applying deep learning to network representation learning in academia.With the help of representation learning in network,the vector representation of node can be used in machine learning tasks on unstructured network graph data,including node classification,link prediction and visualization.Through the analysis and comparison of the existing network representation learning algorithms,it is found that the existing methods have three shortcomings.Firstly,most network representation learning methods only perform representation learning for static networks,ignoring that the network structure would dynamically change over time.Secondly,the existing network representation learning methods do not consider the contribution of network edge attributes or weights to the representation learning.Thirdly,when the structure of network rapidly expands,a large number of new nodes are added to the dynamic network,and the performance of the existing representation learning algorithm might be influenced.In order to solve the above problems,this thesis focuses on the algorithm of dynamic network based on representation learning.The details are as follows:1,This thesis proposes a novel network representation learning algorithm based on probabilistic walks,named Prob Walk.The dynamic network is transformed into a weighted graph,and the probability of choosing the next node during random walk is calculated by the weight value on the edge of the weighted graph,and the library of walking path in weighted graph is constructed for network representation learning.The machine learning task evaluation on the data set proves the effectiveness of the Prob Walk algorithm.2,This thesis proposes a new dynamic network representation learning method.The dynamic network is defined as a sequence of snapshots of static network structure,and use the static network representation learning method to initialize the vector representation of nodes,randomly select a delta vector on the hypersphere,and update the vector representation for the new node with the evolvement of dynamic network.The optimized negative sampling technology is used in representation learning to improve the performance of the algorithm in dynamic network.The experimental results show the effectiveness and efficiency of our algorithm.3,This thesis designs and builds a dynamic network application system to implement the algorithms above to solve the problem of representation learning in dynamic network,and to apply the vector representations of nodes to machine learning applications,such as node classification,link prediction,and network visualization.This system is based on the architecture of Web to support the research of dynamic network.Users can upload customized dynamic network data sets by browser and to perform experiments,including node classification,link prediction and visualization.