Social Media User Modeling Based On Deep Learning

In modern society,online social media has become a popular platform for social interactions,where people are able to acquire information,share opinions and enrich their diary life.Owing to the different service emphases,people tend to immerse themselves in multiple social media platforms and hence fully enjoy various services,which facilitates the information growth pertaining to users themselves.In fact,the massive amount of user information may depict their characterization and daily life from diverse perspectives,which propels researchers to implement the user modeling based on social media.As a matter of fact,several efforts have been dedicated to the user modeling of social media,which can be roughly divided into two categories:single platform-based and multiple platforms-based.Beyond that,in this work,we aim to implement the user modeling from both single platform and multiple platforms angles by investigating the following research tasks:'fine-grained privacy detection' and 'user identity linkage across social media'.In a nutshell,our main work and technological contributions can be summarized in threefold:(1)We present a fine-grained privacy detection network based on textual posts.The proposed network introduces a hierarchical attentive network to distinguish the wordlevel and sentence-level confidences,and hence accurately captures the privacy indicators of each post.Meanwhile,the network incorporates the graph-based semantic correlations among personal aspects as a regularization towards the latent representation learning for posts in the context of fine-grained privacy detection.(2)We present a user identity linkage scheme across social media based on heterogeneous multi-modal posts.The proposed scheme first utilizes deep learning techniques to capture the initial representation.Afterward,the time-aware post correlation modeling is devised to incorporate the temporal correlation among users'distributed posts into user similarity modeling.Essentially,the scheme introduces a time decay factor to capture the pair-wise temporal correlation among posts,which intuitively reflects the influence of their post similarity on the final user similarity modeling.In addition,to adaptively characterize the confidence of different modalities towards the user identity linkage,we propose to integrate the attention mechanism into the user similarity modeling.Thereafter,we utilize the multi-layer perceptron to project the latent similarity distribution into the probability space,and thus identify whether the given user accounts refer to the same identity.(3)To validate the effectiveness of the proposed methods,the paper conducts a lot of experiments to compare with other state-of-the-art baselines.Besides,we also implement ablation experiments and provide the corresponding analysis.As a byproduct,we have released the model,codes and involved parameters to facilitate other researchers.In particular,the comprehensive user identity linkage dataset is also released.