Representation Learning And Information Control On Signed Social Networks

With the continuous upgrading of social network services and mobile devices,social platforms,such as We Chat,Weibo,and Tieba,have been closely connected with people’s daily life.The resulting social network data has also become the focus of many researchers.Through the study of social networks,a lot of valuable knowledge can be obtained,and the development of related theories and technical means can be promoted,which is of great significance from academic,commercial,and social perspectives.Although many researchers have conducted comprehensive and in-depth research on social networks,most current works focus on social networks that only contain simple relationships,such as friend networks,classmate networks,and so on.However,more and more social platforms now allow users to provide feedback including opposing emotional information,such as trust and distrust,likes and dislikes,etc.The network composed of this type of user relationship is called a signed social network.The main research problems of signed social network data mining include network evolution and modeling,graph analysis,representation learning,and information control.For the first two,many scientists and physicists have carried out in-depth research and achieved relatively complete research results.However,there are still many gaps and deficiencies in the current research on signed social network representation learning and information control.These two research questions are the basis of many critical practical applications and have high research value.Therefore,this thesis focuses on signed social network representation learning and information control,trying to analyze the problems from a theoretical perspective and provide solutions.To a certain extent,this thesis is supposed to fill the existing research gaps and inspire follow-up research.The main contents of this thesis are summarized as follows:1.For the representation learning of connected signed social network,the concept of balanced and unbalanced paths and the Markov process are integrated,and a new distance metric that considers the sign information of edges is proposed.It overcomes the defect that the classical Markov process cannot use the information of edge signs.The relevant mathematical properties and analytical solutions of the newly proposed distance metric are deduced,and the theoretical results are also instructive for the generalization of other classical distance metrics.From different perspectives,different node proximity metrics are designed according to the new distance metrics,and node representations are calculated by low-order approximation.The method proposed in this thesis considers the number of paths in the network,the length of the paths,and the latent sentiment information of the paths.Overall,the method can capture the complex global structural information of connected signed social networks.2.For the representation learning of disconnected signed social networks,due to the limitations of connectivity and the different mathematical properties of some matrices,a mainstream and efficient paradigm is to sample through truncated random walks and then optimize parameters through a contrastive learning framework.However,due to the limitation of computational cost,the number of samples that can be used is limited,so the approximate sampling distribution is different from the real sampling distribution,which damages the quality of node representation.In this thesis,a signed proximity summary is designed,which does not rely on explicit sampling and can theoretically fuse information from infinite samples,thus eliminating the impact of sample size limitation on representation quality.The representations are computed by summary matrix factorization,and the representational power of the proposed method is theoretically demonstrated by studying the relationship with a classical contrastive learning framework.In order to further improve the efficiency of the method,two tricks are also designed.One improves the sparseness of the summary matrix and the other one avoids calculating the whole summary matrix.3.For the representation learning of signed social networks containing neutral links,the existing sociological theories and representation learning works ignore or cannot effectively utilize the information of neutral links,thus we design a signed latent factor model inspired by the latent factor model.By introducing positive and negative latent factors,four kinds of node-pair relationships including neutral links are modeled.A loss function based on log-maximum likelihood is designed and an optimization method for latent factor vectors is derived.It is verified by experiments that the information of neutral links can help improve the quality of node representations,and modeling the four types of node-pair relationships is helpful for predicting specific types of links.4.For the information control of signed social networks,considering the sign information contained in edges,we fuse the structural balance theory with the classic information dissemination model and propose an extended information dissemination model.The model can fully consider the complex network structure and network elements.We derive the contribution of the internal information of each node to the total amount of network information.Based on this,the intervention method of the internal information is designed to perform network control and it is proven that the method can achieve optimal solutions.We derive the impact on the total amount of network information by iterative interventions on external information.We then propose a greedy method for information control and design a high-efficiency trick through matrix calculation techniques.