The Sign Prediction Models Based On Transfer Learning In Unlabeled Complex Networks

People's daily social activities are inseparable from the support of a huge social network.Therefore,the data mining work in the social network has become one of the most urgent research topics.However,the sign information in many social networks is often very rare or even completely missing.Due to people's daily behavior habits and huge labor costs,it is very difficult to obtain the label.However,sign information is very important for people's daily life and interpersonal communication.Therefore,in order to obtain the important sign information,transfer learning can be used as an effective method to extract useful information from some networks with known label information,in order to obtain the sign information in the unlabeled networks of interest.Most of the existing sign prediction methods are supervised learning methods,that is,a part of the prior knowledge of the target domain network label is needed to complete the training of the classifier,while a few prediction methods that do not need the target domain label are not accurate enough due to the limitation of the accuracy of the optimal solution.Therefore,this paper improves the exact solution of the optimal solution in the unlabeled network: This paper proposes a sign prediction method based on branch and bound transfer method in the unlabeled network.This method uses the branch and bound method to solve the objective function to obtain the global optimal solution in the complex network space,so that the most useful label information from the network can be mined out,and then assist the target domain to complete the sign prediction.Experiments show that the sign prediction performance of this method is significantly improved compared with the classical methods.In addition,the knowledge transfer method based on deep learning is also a set of theories which are focused on at present.However,this kind of method contains parameters to be updated.In each training,because of the scale of signed network is too large,the gradient change abnormality and calculation amount and other problems,this kind of method is not suitable for signed network.Therefore,this paper proposes a non-deep knowledge transfer method for signed networks: the cross-domain knowledge transfer method based on tri-level program to complete the sign prediction in unlabeled networks.This method integrates the three problems of source domain knowledge extraction,key instances extraction and target domain sign prediction into a whole optimization problem,which greatly reduces the calculation amount in the process of knowledge transfer and avoids too many parameter updating problems.Experiments show that this method can greatly improve the accuracy of sign prediction compared with other methods.When the correlation between the source domain and the target domain is very small or even completely independent,the existing transfer learning methods can not be well applied to the signed network.Because of the huge difference between the source domain and the target domain,the direct knowledge transfer will produce serious negative transfer phenomenon,especially in the signed network.Therefore,in order to reduce the negative transfer phenomenon,this paper proposes a sign prediction method based on three domain relationship model group.The tri-domain relationship model group selects a suitable intermediate domain network as a bridge to complete the knowledge transfer from the source domain network to the target domain network.At the same time,the tri-domain relationship model group also measures the correlation among the source domain,the intermediate domain and the target domain,picks out and removes the interference instances that affect the knowledge transfer,and purify the transferable knowledge in the source domain network.The following experiments verify the good performance of the method in selection for the intermediate domain and sign prediction in target domain.