Key Technique And Application Of Information Fusion Across Social Networks

With the rise of mobile Internet,online social network platforms with different functions has been increasingly developed.To make full use of the platform functions,people are gradually used to participating in multiple social platforms at the same time.This multi-platform behavior and the existence of overlapping users have built a logical connection for the originally isolated platforms.For various reasons,the information between multiple platforms usually lacks explicit association.This sparse explicit relationship leads to the information isolation problem.Aiming at this problem,this paper studies the key technique and application of information fusion across social networks.Based on the cross-platform information fusion,following research could integrate multi-dimensional information in different platforms.At the same time,it could also be used to make a more comprehensive understanding and description of various objects in the platform,so as to provide more possibilities for social network research.Through the analysis of the social network composition and tasks,this paper extracts the key objects of information fusion,i.e.,“users” and “messages”.We first aim at the key technique of “user identity fusion” and study the unsupervised user identity linkage problem,then focus on the key application of “fusion-based message diffusion” and study cross-platform influence maximization problem without definite label.Due to the cross-platform features and the requirement of real environment,related research is facing many challenges.Unsupervised user identity linkage research mainly faces three challenges: how to deal with the heterogeneity of social network information,how to effectively use cross-platform information,and how to deal with the similar behavior of platform users.Cross-platform influence maximization research faces two main challenges: how to model the cross-platform message propagation process in the actual scenario,and how to effectively measure the importance of linkage user in the propagation.In addition,there are still some deficiencies in the existing researches for the two problems: in the former researches,there are deficiencies in the utilization of identity-related information,and the combination of sociology knowledge is lacked? In the latter researches,the real scenario without definite label has not been well studied.The overcome the above challenges and deficiencies,this paper explores the following three research aspects:(1)Aiming at the unsupervised user identity linkage problem,we propose an general unsupervised user identity linkage approach based on retrofitting embedding.It could take any user embedding as input and retrofit it in the embedding space to improve its linkage ability.Our approach first comprehensively utilizes user profile,content and network attributes to extract the user-discriminative features,and constructs cross-platform similar user pairs? then we design the training objective based on similar user pairs to retrofit the user embedding?finally,we calculate the linkage score based on the retrofitted embedding.The experimental results based on real social platform datasets prove that our approach does not rely on certain user embedding method and could effectively improve the cross-platform identity linkage performance.(2)Aiming at the unsupervised user identity linkage problem,we propose an unsupervised user identity linkage framework based on essence representation.According to the sociological theory,the framework introduces the concept of user essence representation in the linkage task,which integrates user’s natural and social attributes,and could obtain the essential differences and identity discriminative information between users.Firstly,the framework adopts user natural attributes to learn the user natural embedding.Then through the processes such as edge confidence generation,user pair set construction and pair-based representation training,it integrates the user social attributes to learn the user essence representation,and finally makes the user identity linkage with essence representation.The experimental results based on real social platform datasets prove the efficient linkage performance of this framework,and get the improved performance combined with the general approach of retrofitting embedding.(3)Based on the user identity linkage technique,aiming at the cross-platform influence maximization problem without definite label,we propose a corresponding diffusion model and two specific algorithms.We firstly define the actual problem,and propose the two-phase diffusion model based on linkage user.Then we further prove the submodularity of the problem under the diffusion model,and design a cross-platform greedy algorithm.We also propose a heuristic algorithm based on linkage user unreachability,which uses network embedding to explore the user unreachability,and combines user influence,together with linkage user,to search the initial seeds.The experimental results based on real social platform datasets prove the efficiency of the proposed algorithm.Finally,according to the research above,we design and implement the prototype system of cross-platform information fusion.The prototype system could provide the function for the above tasks and offer the task results for the users via user-friendly interactive interface.