| In the era of social media,users like to present themselves using both text and pictures.However,picture composition is much difficulty than text composition,therefore,it is reasonable to make presentation by using user generated text and pictures from others which is consistent with the text.There are very large amount of pictures in the social networks,How could users quickly select the suitable pictures from the enormous social medias,it is a challenging problem.In this thesis we started a related study on the Weibo.com and Huaban.com platform,the details are as follows:First,we proposed a visual sentiment analysis architecture based on deep learning.To address the problem of noisy data in the known dataset Sentibank,whose images are collected from the social networks.We propose tag confliction based data refinement strategy to reduce the noisy data.Then we select the resampling strategy to increase the uniformity of dataset.Finally,we improve the framework of deep learning by using the regression loss function.The experiments show that both the dataset refinement algorithm and the improved deep learning model are beneficial.?Second,we propose a cross-platform user profiling method,modeling overlapped user profile by extracting the user's personalized information from the two platforms.Firstly,the user information from the two platforms is modeled respectively.Then the similarity of user models from two platform is computed.Finally,the cross-platform user preference model is obtained by weighted summary.Third,we propose a cross-platform image recommendation scheme for user generated text.First,a primary recommended image list is obtained by the content matching.Then the other recommend list is obtained by the emotion matching.A final recommend list is obtained by combining content consistence,emotion consistence and used profile.The experimental results demonstrate that the proposed scheme is much efficient than random recommendation. |
PDF Full Text Request