Cross Media Content Mining And Search For Security Topics In Social Networks Based On Deep Learning

Social networks supported by massive data affect every social individual in real life through its information transmission method different from traditional news media.In particular,information related to social security and national security topics that are mixed in various media forms of social network information directly affects the direction of public opinion in real society.Focusing on issues related to cross media content mining and search for security topics in social networks based on deep learning,researches about acquisition and processing of social network cross-media content,social network content matching based on reinforcement learning,adversarial social network cross-media content search based on generative learning and related security topic content cross-media search mechanism in social networks are conducted in depth.This research is based on cross media content mining,and finally realizes cross-media security topic content search on social networks.The main contributions and innovations of this dissertation are as follows:(1)Focus on the issue that the characteristics of complex social network data have an impact on the efficiency of social network information search,a cross-media content association analysis algorithm for social network security topics(SSCM)is proposed.The algorithm performs deep semantic learning and search semantic feature association processing on content information such as text media information and image media information respectively.The semantic associations between text and image information content of social networks are explored through the effective mining of cross-media content characteristics of social networks based on self-attention mechanism.A deep topic convolutional neural network-based social network security topic search algorithm(DCNN-CSTRS)is proposed.The algorithm is based on local feature extraction of social network content.It is desiged to match and search constructed social network content representation features through cross-media semantic association analysis for verifying the cross-media semantic association analysis.The effectiveness of the SSCM algorithm and the DCNN-CSTRS algorithm are verified by experiments.The DCNN-CSTRS algorithm improves 3%to 10%compared with selected contrast algorithms on the overall NDCG evaluations,improves 9%to 13%compared with selected contrast algorithms on the overall MAP evaluations,2%to 10%compared with selected contrast algorithms on the overall ERR evaluations.(2)Aiming at solving the problems of existing social network information search methods and traditional information search algorithms that only perform similarity calculation and static similarity matching ranking based on the query content,a content search algorithm for specific topics based on deep reinforcement learning MDPMS is proposed.In MDPMS,social network search process is defined following on reinforcement learning and Markov Decision Process(MDP).A dynamic social network content relevance evaluation process is proposed to evaluate the social network content's semantic features.Matching features mining for social network content is used to select the appropriate actions(select or skip the corresponding content as search result)for social network content to build a list of search results according to the target content.The experimental results based on the Sina Weibo dataset show the effectiveness and superiority of the MDPMS method.The MDPMS algorithm improves 7%to 12%compared with selected contrast algorithms on the overall NDCG evaluations,8%to 10%compared with selected contrast algorithms on the overall MAP evaluations.(3)Aiming at solving the problem of multimedia semantic gap faced by cross-media content information search of social networks,a cross-media information search algorithm based on adversarial generative learning CMSAL is proposed.The CMSAL algorithm is an extension of the social media security topic cross-media content association analysis algorithm SSCM.It combines the social network cross-media data features to generate cross-media feature representations for social network cross media search that follow the or-iginal modal and hashtag distribution.The CMSAL algorithm utilizes the local semantic features of word groups in text and the construction of semantic blocks in images to maximize the correlation between cross-media data based on adversarial learning.The proposed adversarial learning framework is effectively used to rank search results.Experiments verify the effectiveness of the proposed CMSAL for cross-media data sear-ch of full-topic content in social networks.On Sina Weibo dataset,the CMSAL algorithm improves 2%to 11%compared with selected contrast algorithms on the overall MAP evaluations,2%to 10%compared with selected contrast algorithms on the overall precision evaluations for both text search image and image search text tasks.On Wikipedia dataset,the CMSAL algorithm improves 2%to 8%compared with selected contrast algorithms on the overall MAP evaluations,2%to 10%compared with selected contrast algorithms on the overall precision evaluations for both text search image and image search text tasks.On NUSWIDE dataset,the CMSAL algorithm improves 3%to 7%compared with selected contrast algorithms on the overall MAP evaluations,3%to 5%compared with selected contrast algorithms on the overall precision evaluations for both text search image and image search text tasks.(4)Combining the proposed algorithms of SSCM,DCNN-CSTRS,MDPMS and CMSAL,a deep learning based social network cross-media content search system is implemented.The system includes social network cross-media content acquisition and processing module,social network reinforcement learning content search module,generative adversarial learning-based social media cross-media content search module and social Network Security Topic Content Search Module.The implemented system reliably evaluates and displays the search results of the proposed methods from the perspective of data visualization.The effectiveness of several algorithms proposed in this paper is verified by this system.