Recommending Social Media Topic By Using Network Representation And User Content

In recent years,we have seen the rapid development of social medias,where more and more users are participating in.With the prevalence and popularity of topic participation,understanding which topics a user is interested in has become an emerging research problem,which can greatly benefit many downstream business applications including commodity recommendation,online precise targeting,customer retention management,even disease spread detection and public health monitoring,etc.Meanwhile,existing efforts on topic participation forecasting mainly focus on learning from historical user-generated texts to infer their preferred topics,or simply recommending topics that user's friends are interested in,which can't fully consider the user's feature,they usually lead to a poor recommendation performance.In this work,we first collect the data from four major social media platforms,including Weibo,Zhihu,Douban and Twitter.We use a classic topic modeling(LDA)to quantify both intrinsic and extrinsic influence on topic distribution.The results shows that majority of users have consistent topic interest with their past and have similar preference to their friends.These findings motivate us to model user historical text and their social network when recommending which topics a user is likely to join in the future.We formally model it as a multi-label classification where posting content and social network structures are jointly learned to characterize latent relationships between topic participation and users.Our proposed method consists of two major components:(1)user text embedding;and(2)network embedding – combined to learn user intrinsic and extrinsic preferences.To facilitate effective learning,we further propose a novel framework that converts continuous latent embedding vectors into binary representations via Locality-Sensitive Hashing(LSH),which preserves both content and network similarities that reflect user intrinsic and extrinsic preferences.Such a design via LSH transformation can improve the efficiency of convolution operations – more importantly,it can well distinguish users in the low-dimensional space while preserving their similarities,significantly enhancing the recommendation performance.To explain the behavior of the proposed model,we use the influence function to interpret its model ability by investigating the importance of each training sample on the recommend performance of every testing instance.We conduct extensive experiments on the collected datasets and the experimental results demonstrate that our method can significantly boost the recommend accuracy in terms of all metrics,as well as the model efficiency,compared to existing baselines,which fully demonstrates the effectiveness of the proposed model.