Label Inference Research Based On Maximal Entropy Random Walk Over Social Network

With the rapid development of the Internet,social networking platform has developed rapidly.With the increase of the large-scale entities and the large-scale of relationships of the entities in the social network,graph structure is widely used in dealing with the individuals and the complex relationships of individuals.Individuals in social networks have different labels,however,in social networks,the labels are always labeled by the users in the social network and because of the privacy protection,many users choose not to label themselves in the social network.Therefore,the labels in the social network are always incomplete.In real-world applications,the complete label system is necessary.such as personalized services and merchandising etc.,accurate label reasoning algorithm is very important and will bring large profit for us.Also,it can support more convenient for users.In this paper,we propose a method which is based in the structure of the graphs to solve the problem in the social network label inference.The main research results and innovation point embodied in the following aspects:1.A new label reasoning problem is proposed,because of the problem of less open datasets,the DBLP data migration to the current task,which can be help us to build the model ready.2.Analysis the dataset and the research task,process the statistical data,and then apply the dataset to the algorithm.3.A new label reasoning model based on maximum entropy is proposed.First,the random walk model is used to the entire social network structure and to find the structure similar of the vertices.At the same time in the end of each step of the random walk is completed and the maximum entropy is used to make the algorithm more accurate and convergent faster.4.According to apply the method in single-label graph and multi-label graph,we can find that this method can achieve good result.Also,it can help us to revise the uncorrect result by research the relationship of the labels in multi-label graph.5.According to the proposed algorithm,the corresponding time complexity analysis and space complexity analysis is proposed,we can see that our algorithm in time and space efficiency is relatively high contrast to other corresponding algorithm.The proposed algorithms are applied to the actual data set,compared to other methods,have achieved a ratio better effect.And from the experiment we can see that a better effect when the amount of the unlabeled vertices is large,our algorithm still maintains a high accuracy rate.