Feature Selection Mechanism For Multimodal Social Media Data With Privacy Protection

In recent years,social network has become an important part of people's life.Compared with traditional news reports,social network platforms encourage users to produce their own content.News is more real-time and diversified.The content often integrates multiple modes such as text,pictures and videos.People can get news more quickly and conveniently through social networks.At the same time,the quality of users is uneven,the objectivity of the news can not be guaranteed,therefore fake news spreads rapidly in social networks,misleading the masses,destroying the stability of social.So the detection of fake news in social networks is particularly important.This paper studies the feature description in multimodal data mining of Twitter and microblog,focuses on the importance evaluation of multimodal feature and feature selection of fake news detection on social platforms,as well as the research on privacy protection and the utility of data.Multimodality brings feature dimension explosion.The features contains a lot of noise,which not only affects the training time and hardware consumption,but also affects the detection accuracy.For solving this problem,this thesis combines a variety of feature selection methods to optimize group features.Aiming at the feature optimization mechanism,this thesis mainly carries out two stages of work: 1)combining principal component analysis,its improved method and recursive feature elimination method to optimize multi-modal feature set.2)Multimodal features mean that features exist independently in groups.It is difficult to evaluate the importance of group features by traditional feature selection methods.In this thesis,multi-kernel learning method is used to evaluate multimodal group features and find the weight of group features,so as to realize the optimization of group features.Combining the two points above,the FS-MKL group feature selection method is designed and implemented.Based on Twitter and microblog data sets,the performance of FS-MKL is compared with other detection models,and the effectiveness of FS-MKL method is verified.Social network data contains a lot of user information such as locations,pictures,social relations(the number of followers and fans.)and etc.,once used maliciously,the results are very serious.In order to protect the privacy of users,this thesis designs and implements a new data publishing algorithm SDRRDP,and proves that SDRRDP algorithm can effectively improve the availability of data while protecting the security of data.Experimental results show that compared with other disturbance algorithms,using the disturbance data published by SDRRDP for model training can get better training effect.This proves that the combination of SDRRDP and FS-MKL solves the problem of user privacy security.Finally,in order to improve FS-MKL,promote the performance of feature selection algorithm,optimize the parameter setting of group feature selection,and finally improve the performance of false news detection,particle swarm optimization algorithm is adopted to improve FS-MKL,and PSO-FS-MKL group feature selection mechanism is formed.Experiments show that PSO-FS-MKL is better than FS-MKL in time performance and accuracy.