Reserch And Implement Of User Context Information Perception Algorithm Based On The Social Pictures

As the networking of social behavior and the multi-media information of the mobile Internet is growing exponentially,especially represented by social images,which,to a better extent,matches with the characteristics of time fragmentation,behavior movement and ultimate experience in the age of mobile Internet.The analysis of social images that were created by users can be used to research on the user context information,which is used to describe the users themselves and any related information and has great significance in recommender system and personalization.It mainly adopt the method of information extraction,by which the information is extracted from user registration messages or historical behaviors.However,there is a considerable limitation in this method due to privacy protection.To solve this problem,this work proposes a novel and effective perception method to predict user context information by using social pictures.This work starts with the social image of the mobile Internet and users who post them.Firstly,two experimental datasets are acquired(from Wechat and Sina microblog)and the user context information is defined,such as user gender,the posting activity,the geographical location of photos,the terminal,the posting habit,the social influence and the images' impact.Secondly,this work analyzes the relationship between users' information and low-level features extracted from users' images.Next,the work proposed a user characteristic learning method based on bag of visual word(UCL-BOVW).Then models are built for prediction based on the mentioned two data characteristic: a model based on image low-level feature fusion,and an information prediction model based on user characteristic learning.Finally,prediction is made for all user context with the mentioned two models combining with machine learning algorithm.Additionally,this thesis also optimizes the model by leveraging the other user information in the collected data.The experiment result demonstrates that the proposed method and GBDT algorithm can effectively predict the user context such as user's image influence,activity and so on.