Link Prediction In Dynamic Networks Based On Evolution Of Similarity

Link prediction is an important method to mine potential relationships of data.Most networks in real life are the dynamic networks,but the traditional link prediction algorithms assume that the networks are static.Thus,their performance is restricted.Based on the traditional link prediction algorithm,this thesis uses the structural information and temporal information of networks to solve the two important problems in dynamic network link prediction research:1.measure of the similarity between nodes in the static network,2.forecasting of the similarity in the future.We propose a similarity index based on hybrid structural information and a reward forecasting model,and finally obtained a hybrid structural reward prediction algorithm for the dynamic networks.The detailed contents of this dissertation are summarized as follows.1.We propose a hybrid structural similarity index in static networks.In order to make full use of the evolutionary process and temporal information of the networks,we divide the integral network into sub-networks according to historical time,and establish the time series of the network.We use the process of the change of similarity between nodes in the sub-networks to describe the evolution of the network.Because the sub-networks are all static networks,the similarity index based on the structural information of networks is used to measure the similarity between nodes in the sub-network.Based on the local structural information,we introduce the link information between common neighbors and propose a hybrid structural similarity index that is suitable for static networks.2.We propose a hybrid structural linear regression algorithm in dynamic networks.In this thesis,based on the hybrid structural similarity index applied to the static network,we combine the hybrid structural similarity index with the linear regression model,and obtain a hybrid structural linear regression algorithm suitable for dynamic networks.This algorithm establishes the sub-network similarity score time series between nodes by using the hybrid structural similarity index,uses the linear regression model to predict the future similarity between nodes,and calculates the final similarity score to complete the link prediction in dynamic networks.3.We propose a hybrid structural reward prediction algorithm in dynamic networks.In link prediction research,the forecasting model directly affects the final performance of the algorithm.In this thesis,we study the forecasting model based on the hybrid structural linear regression algorithm and propose a reward forecasting model.We combine this model and the hybrid structural similarity index and obtain the hybrid structural reward prediction algorithm.This algorithm makes full use of the structural characteristics and evolutionary process of networks,and has higher prediction precision than traditional static algorithms and dynamic algorithms in the real network data sets.