Online Social Network Multi-Modal Security Content Search Based On Deep Reinforcement Learning

The rapidly developing social network has generated a large amount of data and it also generates considerable incremental data in real time.On the one hand,the unique data characteristics of social network big data make traditional information search algorithms difficult to meet the needs of users.On the other hand,the data diversity of social network information puts higher demands on social network big data acquisition and processing.Improving the accuracy and efficiency of related search algorithms have become research hotspots due to the phenomenon.The main works are as follows:(1)Social network data perception and extraction algorithm based on content filtering based on deep convolutional neural network is proposed to perceive and extract security content from microblogs.The content about the security topics is defined as the query.The collection of the microblog document is defined as the candidate set of the microblog search.According to the text and image features of microblog contents,the evaluation among microblog dependency is obtained.(2)Social network multi-modal content mapping and recognition based on deep learning algorithm(M-RSCNN)is proposed.The algorithm combines the image information contained in the microblog content with the modified eigenvalues to obtain a consistent semantic space.The microblog short text and the microblog document set composed of image information,time information and spatial information are input into the trained convolutional neural network model.Corresponding semantic space is obtained by nonlinear transformation such as convolution calculation and pooling calculation.A common semantic space is established based on the convolution features,in which the representations in the space are the results with mapping and recognition.(3)A microblog text content matching and search method based on reinforcement learning algorithm is proposed(RSDQN).Microblog search is re-defined following the concept of reinforcement learning such as action and reward.Matching scores of the individual document of the microblog are calculated by the reinforcement learning network model and are regulated by the image modal information.According to the relative value of the scores.The current search result will be determined to join the overall search results set or not.Normalized Discounted Cumulative Gain(NDCG)of overall search results is calculated in real time.The obtained single result score and the overall search result NDCG are re-input to the enhanced network generation model as double feedback for repeated learning.The self-iteration of the search network is implemented to complete the matching,sorting and searching.(4)A multi-modal security content search system for online social networks is developed.The system consists of three modules:the multi-modal big data perception and extraction in online social network,the multi-modal content mapping and recognition of social networks based on deep learning and the social network content matching,sorting and searching based on reinforcement learning.The system is fully functional and user-friendly.