| With the rapid development of the Mobile Internet,social networks are booming in recent years.Community detection,influence maximization,and personalized recommendation emerged one after another.As an important basis of above studies,user profiles can be divided into two parts:user static profiles that describe users’ static attributes such as gender and age,and user dynamic profiles that characterize users’dynamic attributes such as interests and hobbies.Owing to its importance for many applications,the user profiling problem has been widely studied.Through systematic research and analysis,this thesis found that most of the existing studies focus on mining information from a single social network.However,people often join multiple social networks for different services.This phenomenon leads to both correlations and great differences in user information from different social networks.Therefore,user profiles,established on information from a single social network,often suffer from inaccuracies.It is necessary to integrate information from different social networks for more accurate user profiling.Thus,this thesis studies three key problems,aiming at establishing more accurate user profiles with information from different social networks.Firstly,this thesis studies the social network user alignment problem to obtain user accounts of the same person in different social networks.Then,based on user alignment information,the multi-social network user static profiling problem and the multi-social network user dynamic profiling problem are studied respectively.The main contributions of this thesis are introduced as follows.1.For the problem of user alignment,this thesis proposes a user alignment method based on behavior analysis,called BANANA-RGB.Most of the existing studies focus on utilizing similarities of users’ static information(e.g.,structure or attribute information),which lacks consideration of behavior information for promoting user alignment.To address the above limitation,this thesis proposes BANANA-RGB based on behavior analysis.Firstly,to capture users’ multi-scale behavior information,this thesis trains a variant of the hierarchical periodic memory network with personalized memorization.Then,this thesis designs a tensor fusion network-based alignment component to fuse users’ behavior and static information,and construct user pairs.It fuses representations of user pairs and calculates alignment probabilities in multiple subspaces.Finally,based on the alignment information,this thesis proposes a gating-based cross-network behavior fusion component to integrate correlated behavior information from the source social network for facilitating the predictive analysis of behaviors in the target social network.Extensive experiments on real-world datasets demonstrate that BANANA-RGB outperforms state-of-the-art methods in user alignment and behavior prediction.2.For the problem of user static profiling,this thesis proposes a user static profiling method based on multi-social network fusion,named HAMLET.Most of the previous works either focus on mining information from a single social network,or ignore correlations between profiling tasks.To address the above limitations,this thesis proposes HAMLET by fusing information from different social networks.Firstly,this thesis proposes a hierarchical attention-based network that assigns various features with different attention weights,and fuses these features for single-network user representations.It fuses user representations from different social networks,based on the attention mechanism,for learning multi-social network user representations.Then,this thesis proposes a multi-task self-training algorithm.Based on the correlations,the algorithm alleviates the missing problem of some user attribute labels and improves the performance of user static profiling tasks.This thesis conducts extensive experiments on realworld datasets,and experimental results demonstrate the substantial superiority of HAMLET against state-of-the-art methods.3.For the problem of user dynamic profiling,this thesis proposes a user dynamic profiling method based on multi-social network fusion,named HABIT.Most of the previous studies focus on mining user information in a single social network.To address the above problems,this thesis proposes HABIT based on multi-social network fusion.Firstly,this thesis proposes a single network user representation component,including a hypergraph attention network and a Transformer encoder.The hypergraph attention network aims to mine high-order relations between keywords.And the Transformer encoder models dependencies among usergenerated content.Then,this thesis proposes an information fusion component.This component introduces several fusion units to facilitate the information interaction between user representations,from different social networks.The information fusion component learns user-specific label representations based on the attention mechanism and general label representations with a feedforward neural network.Finally,this thesis calculates the probability of whether a user has a specific tag,based on the two tag representations.Extensive experiments on two real-world datasets demonstrate the superiority of HABIT against state-of-the-art methods. |
