Research On User Profile Modeling Based On Multi-modal Social Media Data

With the rapid development of Internet,social media has entered people's perspectives,and has influenced and changed people's daily lives.Social media is a public,open platform and tool,where people can freely communicate with others,express themselves and share hot topics,thus producing massive amounts of data.How to make use of the massive data of social media platforms to provide users with valuable services is the focus of current research.The user profile is a"tagged"user model,abstracted by the information of user's basic attribute contents and behavior contents.The research results are of great significance for intelligent recommendation,digital marketing,information retrieval and so on.This thesis analyzes and constructs the user profile model based on the data from social media,focusing on the uni-modal dynamic user profile and multi-modal user profile.This thesis proposes a novel Dynamic User Profile Model?DUPM?,where the users'attributes are finally constructed through the acquisition of social media data,text preprocessing,topic extraction,and user attribute mining.In particular,we introduce the concept of time windows to add time factor and provide the attenuation function to obtain dynamic user attributes,which helps to predict and analyze users'characteristic in the near future.Extensive experiments are performed to verify the effectiveness of our model.In addition,a multi-modal User Profile Model?mmUPM?is proposed in the thesis.The model construction is realized through the research on multi-modal data acquisition,text and image feature extraction and fusion methods,and user profile deep learning algorithm.In the process of implementation,we uses the Poisson Gamma Belief Network?PGBN?proposed by Zhou^[29]to construct a five-layer deep learning network,which is suitable for large-scale social media platform data.Finally,we utilize four multi-label algorithms for experimental evaluation and a large number of comparative experiments have proved the feasibility of the proposed model.