Topic Popularity Prediction On Complex Social Networks With Deep Learning And Grey Methods

With the scale of social network users continuing to expand,the scale of social network users’ participation in topic discussion has increased.The topics with high attention,high degree of opinion recognition,clear and influential opinions have gradually become the sprout of network public opinion,and have a very important and far-reaching impact on the formation of Internet users’ views and even social views.By predicting the popularity of social network topics,people can allocate resources in advance,arrange work reasonably,prevent adverse situations and control the dissemination of information.The topic popularity sequence of complex social network has the characteristics of diversity,nonlinearity,small sample and dynamic.The essence of topic popularity prediction in complex social network is the temporal feature mining problem of topic popularity time series.In recent years,the excellent performance of deep learning in autonomous learning features of time series has been widely used.In addition,due to the diversity of influencing factors of social network topics,topic popularity prediction is often of uncertain,and the grey methods have great advantages in dealing with uncertain data.Therefore,it is urgent to study the topic popularity prediction based on deep learning and grey methods.This thesis concerns about topic popularity prediction on social networks under multi-topic environment,and conducts a series of studies and explorations based on deep learning and grey methods.It mainly includes the following aspects.(1)This thesis studies the quantification and calculation method of topic popularity,and puts forward it under the background of multiple topics based on the degree of grey incidence.The weight of topic popularity index is calculated by using analytic hierarchy process(AHP).At the same time,the effectiveness of the quantitative method of topic popularity relevance is proved by the visual heat map of grey correlation.(2)This thesis studies the prediction problem of topic community popularity,proposes a topic community popularity detection method based on local nearest neighbor propagation and random walk algorithm,and uses recurrent neural network(RNN)to predict topic community popularity.According to the local neighborhood proximity and random walk probability,the topic community popularity is calculated,and the effectiveness of the model is proved by experiments.(3)In this thesis,the prediction problem of non-burst topic popularity is studied,and a prediction model and algorithm with collaborative training based on GM(1,1)method and RNN network is proposed.Then,the prediction results are expressed by Markov probability transfer matrix,which can accurately capture the temporal evolution characteristics of non-burst topics,and applied to real social network platform.(4)This thesis studies the problem of burst topic prediction,and proposes a prediction model and method based on the degree of grey incidence and LSTM network.The grey correlation feature is extracted from the time series of burst topics,and a prediction model integrating GM(1,1)and LSTM is established.The effectiveness of this improved model is verified on real data sets.