| Online social network is an online platform that allows people to establish social connections,communicate with each other,and share experiences.With the vigorous development of Internet information technology,online social networks are favored by users for their instant communication,convenience,diverse information,and massive users.People extend daily social behaviors from offline life to online networks,making online social networks an indispensable and important platform in people's daily lives.As information producers and consumers in online social networks,users are the core assets of various online social platforms.Their social characteristics are important information to identify users,describe user characteristics,portray user portraits,and understand user intentions.Various online social network services analyze user's social characteristic information,mining user identity,user composition,user tendency,providing different services like business intelligence analysis,political situation analysis,personalized recommendation service,real-time medical analysis.Therefore,the analysis of user's social characteristics has practical significance and application value.However,there exists many challenges in social characteristics analysis.On the one hand,the number of online social network users is extremely huge,ranging from tens of millions to hundreds of millions,and the user attribute information has diverse categories and types.The increase of the size of users and attributes will cause output space to increase exponentially,greatly increasing the difficulty of user attribute inference in terms of accuracy and timeliness.On the other hand,the relationship structure between online social network users is complex,heterogeneous and diverse,and the semantic information of the relationship between users is implicit and scarce.How to effectively use low-quality training data to analyze social relationship has become a challenging task.Faced with these practical challenges,this paper studies user attributes from four perspectives of user attribute information and user relationship information,respectively from user explicit attribute information,user implicit attribute information,user implicit social relationship and user relationship type.There are four items:information mining,points-of-interest recommendation,hidden social relationship discovery between users,and relationship type identification between users.(1)In terms of user attribute information mining,due to the fact that some user attributes are missing or difficult to distinguish,this paper propose a user attribute inference framework based on social relations and user generated content,describing the probabilities of user attribute separately.At the level of social relations,this paper proposes an attribute confidence propagation algorithm SocialInfer.According to the phenomenon that users with social relationships in real society have similar attributes,the algorithm uses confidence spread to infer user attributes.At the level of user generated content,this paper proposes a UCBert model.By combining the characteristics of different user groups in terms of language environment,language characteristics and language habits,the model uses a two-way Transformer model to infer the user's personal attributes.By synthesizing the influence of these two different factors,this framework calculates the probability distribution of user attributes and mines the attribute with the highest probability.The experimental results are tested and verified on Facebook,Twitter,and Instagram,proving that the proposed method has higher accuracy of attribute mining.(2)In terms of points-of-interest recommendation,this paper proposes a unified framework for the challenges of sparse user-interest matrix and mixed context information.We unify the influence from user preferences,friend importance and position of interest points.For the user preference,the model uses collaborative filtering to calculate the similarity of each user's check-in trajectory.For the friend importance evaluation,the model uses a bookmark coloring algorithm to calculate the user's relevance to POIs.For the relevance of POI location,we assume the distance distribution between user interest points and estimate candidate POIs according to distances of pairs of POIs.Finally,we summarize the prediction score of by three factors to recommend potential POIs.This paper finally conducts experimental tests on real datasets,which can have better performance in terms of accuracy and recall.(3)In terms of implicit social relationship discovery,in view of the problems of incomplete network structure and node information,we propose an implicit relationship discovery model based on social augment graph.The model builds an augment graph to add more information to hidden nodes by fusing a small amount of visible social relationships,user attribute information,and user generated content.We use convolutional neural networks to convolve different types of relationships on the augmented graph to get the final node embeddings.Then,we infer possible relationships between nodes.The experimental results show that our proposed model shows excellent performance on various datasets.(4)In terms of relationship type recognition,we propose a relationship recognition model based on graph attention mechanism to solve the problem of heterogeneous interconnection of social relationships on OSNs.By analyzing the cutting edges in the networks,the graph attention mechanism is used to learn two different edge types.Sociological theories such as structural balance theory and state theory point out that the relationship structure within a specific triple is balanced.By distinguishing the triple structure formed by different types of edges,the graph attention mechanism is used to learn the characteristics of its network structure.Finally,by constructing a suitable objective function,we obtain the final node embeddings to infer edge types.The test results of multiple data sets show that this method has higher accuracy. |
PDF Full Text Request