Research On Social Network Alignment

With the development of the Internet,these years have witnessed the booming of social networks.Nowadays,social network is becoming an indispensable part of people's everyday life.People usually join in multiple social networks to enjoy more diverse services.People are in habit of delivering instant messages on WeChat,sharing college activities on RenRen,publishing daily status on Weibo and networking professional occupations on LinkedIn.Actually,mining the user correspondence across social networks is of fundamental importance.This thesis formally terms this problem as social network alignment.In terms of commerce,social network alignment paves the way for a series of downstream applications,such as user profiling,information diffusion,cross-domain recommend-ation and multi-platform event organization.In terms of national security,social network alignment constructs the foundation for several critical tasks,such as user identity linkage,anonymous user identification and public opinion analysis.Thus,the social network alignment is receiving an increasing attention from academic society.Existing studies focus on social network alignment at user-level,and show limitations in aligning multiple static social networks with robustness and aligning dynamic social networks.In social networks,more often than not,users tend to organize into communities.With social network itself regarding as a macroscopic unity,users contribute to the microscopic view while communities contribute to the mesoscopic view.However,alignment at community-level has been rarely touched before.To this end,this thesis systematically studies on the social network alignment at user-level as well as community-level.The user-level alignment aims at analyzing the natural individual while the community-level alignment aims at analyzing the real-world groups.For user-level alignment,according to the extent of structure evolvement over time,the thesis considers two scenarios,i.e.,user-level alignment across static social networks and user-level alignment across dynamic social networks.For community-level alignment,the thesis focuses on analyzing the real-world group with a relatively stable organization in the long-term observation,and thereby studies the problem of community-level alignment across social networks regardless of structure evolvement.The main contributions of this thesis are listed as follows:(1)For the problem of user-level alignment across static social networks,this thesis proposes to,across multiple social networks,integrate attribute and structure embedding for reconciliation,referred to as MASTER.In the literature,there exists a series of prior studies on user alignment between social networks.This study summarizes the limitations of prior studies in two aspects,i.e.,multiplicity and robustness.To address these limitations,MASTER leverages both attribute and structure information and align multiple social networks in a common subspace robustly,which is formulated as an optimization problem of a constrained collaborative matrix factorization.To address the optimization problem,a nonconvex splitting alternating algorithm is proposed,and further theoretical analysis is given.To further improve the efficiency of MASTER,a novel clustering approach is proposed so that MASTER can be performed in each cluster in parallel.The clustering approach is a unity of two steps,i.e.,augmented pre-embedding and balance-aware fuzzy clustering,to guarantee the high accuracy and high efficiency.Each step is formulated as an optimization problem and a cost-effective optimizing algorithm is proposed.Finally,extensive experiments on several real-world datasets demonstrate the substantial superiority of the proposed approaches against the state-of-the-art methods in terms of precision.(2)For the problem of user-level alignment across dynamic social networks,this thesis proposes a neural network-based approach,Dynamic Graph autoencoder based dynamic social network Alignment,referred to as DGA.Dynamics is an inherent characteristic of some social networks,such as the online friending platforms.However,existing studies on user alignment assume the social networks to be static,which cannot cope with dynamic social network alignment.To bridge this gap,this thesis proposes to study the problem of dynamic social network alignment.It faces the challenges in inter-network dynamics modeling,inter-network alignment modeling and optimization.In DGA,to model the inter-network dynamics,a novel neural network architecture of dynamic graph autoencoder is proposed.To model inter-network alignment,a common subspace is constructed via semi-nonnegative matrix factorization for aligning users.An optimization framework is formulated to collaborate the neural network and matrix factorization.To address the optimization problem,an effective and theoretically grounded alternating algorithm is proposed.In particular,the multiplicative updating rule is derived for semi-nonnegative matrix factorization,and further theoretical analysis is given.Finally,extensive experiments on several real-world datasets demonstrate the substantial superiority of the proposed approach against the state-of-the-art methods in terms of precision.(3)For the problem of community-level alignment across social networks,this thesis proposes a unified hyperbolic embedding approach for embedding and aligning community,referred to as PERPECT.In social networks,the user provides its microscopic perspective while the community provides its mesoscopic perspective.In the literature,user alignment is extensively studied,however,community alignment is rarely touched.To bridge this gap,this thesis proposes to study the problem of community alignment across social networks.It faces the challenges in the choice of representation space,the model of community alignment and optimization.Regarding the choice of representation space,it is concluded that hyperbolic space performs more favorable against the Euclidean space,and thus the social network is embedded in hyperbolic space.To model community alignment,the hyperbolic community representation is learned via a proposed hyperbolic mixture model.The hyperbolic common subspace is then constructed via transferring anchor user representation for aligning communities.Formally,an optimization problem with positive semidefinite constraint is formulated.To address the optimization problem,an alternating algorithm based on the Riemannian geometry is proposed,and further theoretical analysis is given.Finally,extensive experiments on several real-world datasets demonstrate the proposed approach aligns the communities across social networks with high accuracy and high quality.