Prediction Of User Relationship In Social Networks Based On Decision Tree Ensembles

Mobile social networks have become a crucial platform for modern information exchange.User relationships play an essential role in the formation of social networks and the assessment of information quality.However,due to problems such as privacy settings or data loss,the relationship between nodes is not visible in many cases,which poses a significant challenge to the integrity and availability of the network.User relationships of different strengths in social networks can be mapped to edge weights in directed weighted networks.Therefore,constructing a high-precision and robust edge weight prediction mechanism in the network is an urgent research topic.At present,some researchers have proposed some methods that use the combination of node attributes and topological features to predict the edge weight in the network.But in practical applications,the properties of nodes are difficult to obtain.Therefore,this dissertation proposes a Directed Edge Weight Prediction model(DEWP)based on the decision tree ensembles model.This algorithm only uses the network topology and does not depend on the private attribute information of the nodes.According to the local similarity index,combined with the factors of direction and weight in the network,a series of indices for measuring the similarity of node pairs in a directed weighted network are designed and used as the characteristic attributes of the edges corresponding to the node pairs.Random forest,Gradient Boosting Decision Tree,XGBoost,LightGBM algorithms are used to train directed edge weight prediction model.In order to prevent overfitting,this dissertation uses the blended model of the above four regression models.In this dissertation,the DEWP marginal weight prediction effect is evaluated on different data sets.Compared with other existing correlation methods,the DEWP predicted edge weights have a smaller root mean square error and a larger Pearson correlation coefficient between the actual edge weights,which shows that DEWP can predict the direction more accurately.The edge weights of the weighted network can predict the user relationship in the social network more accurately.At the same time,this dissertation tests that DEWP has good robustness in a network that lacks different degrees of border weight information.