Research On Link Prediction Method Based On Positive And Unlabeled Examples

Link prediction is aim to predict missing links based on a known structure in the network.Link prediction,as an important research content of network data analysis,has important research significance in exploring network evolution laws and data completion,it also has important application value in the fields of recommendation systems,bioinformatics,and scientific research cooperation.From the perspective of supervised learning,link prediction is regarded as a binary classification problem.The node pairs are regard as examples,the node pair with link is positive example,and the node pair without link is negative example.Link prediction problem is solved by constructing a link prediction classifier.Due to the large-scale and uncertainty of the network,a large number of node pairs that cannot observe links should be regarded as unlabeled examples.Therefore,how to select reliable negative examples from a large number of unlabeled samples becomes the difficulty of constructing a link prediction classifier.At the same time,the traditional classification models are based on the assumption that the examples and the population are independent and identically distributed,but the examples in the network do not satisfy this assumption.This paper regards the node pairs that can be observed the link in the network as positive examples and the node pairs that cannot be observed the link as unlabeled examples.It studies the problem of link prediction based on positive and unlabeled examples(positive and unlabeled examples learning is referred to as PU learning for short).There are some problems in the link prediction based on positive and unlabeled examples,such as the lack of negative examples,and the examples and population isn't independent and identically distributed.In view of the above problems,this paper puts forward effective solutions.The main research results are as follows:First,this paper proposes a link prediction method based on PU learning.This method uses the community structure information and integration method of the network to solve the problem of how to select reliable negative examples from a large number of unlabeled examples to construct a link prediction classifier.By comparing with the existing methods of selecting reliable negative examples on real data sets,experiments show that the reliable negative examples selected by this method can construct a link prediction classifier with better prediction effect.Secondly,this paper proposes a link prediction method based on PU learning andrandom walk.Link prediction classifiers constructed by link prediction methods based on PU learning are all based on the assumption that examples satisfy independent and identical distributions,while examples in the network do not satisfy it.Aiming at the above problems,this paper uses the link prediction classifier and the community structure information of the network to improve the superposed random walk,and proposes a new link prediction method.Experiments show that from the overall prediction results,this method can further improve the accuracy of link prediction.In conclusion,this paper proposes corresponding solutions to the problems in link prediction based on positive and unlabeled examples,improves the accuracy of link prediction,and provides new solutions and ideas for the problem of link prediction based on positive and unlabeled examples.