Social Network Cross-media Big Data Search Based On Deep Learning

The social network site represented by Weibo contains a huge number of valuable national security incident information,and its semantic information exists not only in texts but also in images and user characteristics.Massive cross-media information brings challenges to traditional data warehouse and data retrieval techniques.The deep neural network has emerged in recent years,which provides a new solution for the understanding and extraction of cross-media semantic information.In order to make more effective use of cross-media information for social network national security events,this thesis designs a system to collect and retrieve massive data.The retrieval system adopts the popular search engine architecture.From the perspectives of image features,text features and user features,the deep learning method is used to analyze the social network national security incident data to achieve accurate and efficient retrieval.The main work done in this thesis is as follows:(1)A real time collection method of social network cross-media information and feature extraction based on deep learning are proposed to extract semantic features of social network cross-media.A real time collection strategy for cross-media information in social networks is proposed.Through the study of deep learning theory and common models,deep learning methods are used to extract semantic information in social network images and text information and perform social network cross-media information processing.(2)A classification method based on deep neural network is proposed to achieve effective filtering of social network images and texts which are unrelated to national security.The feature vector is constructed from the user features,explicit content features and hidden content features of Weibo.Combined with the deep learning method,the text and image information are comprehensive analyzed to effectively identify unrelated information.Then we verified the effectiveness of proposed method by conducting text classification experiments on Weibo dataset.(3)A search method based on cross-media text expansion is proposed to achieve the semantic expansion of social network cross-media information.The word embedding representation is used to model word similarity to represent the captured semantics and vector algebra.We use a neuro-language model to learn the word-word association within the microblog text,to mine the potential semantic information of the microblog.We make full use of image information and the co-occurrence relationship of cross-media information to train the semantic mapping between images and texts.Then we verify the proposed method by conducting cross-media semantic expansion and search experiments on Weibo dataset.(4)By combining the research contents of the above three aspects,this thesis designs a deep learning-based social network cross-media big data search system,and then develops the system and implements the proposed algorithms.The system consists of three functional modules:real time social network cross-media information collection module,national security incidents cross-media feature extraction and search module,and social network content and user feature analysis module.