Research On The Method Of Forecasting Binary Asymmetry In Social Networks

The prediction of relationship in social networks provides great support for managers and decision-makers. The arrival of the era of big data provides new ideas and methods for the prediction of the relationship in social networks.In this paper, a method of relationship-prediction based on BP Artificial Neural Network was built in order to solve the problems of limitation of assumptions and the inadequacy of the use of information which exist in normal methods of relationship-prediction. The theory of set and relationship, the theory of complex network and data mining technology were used. In the perspective of the relationship between nodes the evolution and development of the binary asymmetric relationship in social networks was predicted, using the structure information in networks.By constructing a forecasting method based on BP artificial neural network surrounding the problem of the prediction of the binary asymmetric relationship in social networks. the research work carried out as follows:(1) The premise which is needed in the prediction of asymmetric binary relationship based on ANN——the model input and output is not independent. Two methods to test whether the premise is true in a specific network were given. One is based on chi-square test and the other is based on the test of the normality of out degree distribution. Combining with Lyapunov central limit theorem, this paper proved:as long as the out-degree of the network is not normally distributed, the binary relationship between each local node of the network interdependencies, thus this method is suitable for the prediction of binary asymmetric relationship in the network.(2) In order to reduce the amount of calculation, a method of data dimension reduction was designed. And this paper proved that this method of data dimension reduction would not cause loss of information.(3) The positive sample class empowerment was used to solve the problem of imbalance which may exist in the data.(4) The optimal weights using of the neural network was computed, using the steepest descent method. And several initialization method was used to reduce the risk of error due to the local optimization problem.(5) The total prediction accuracy and positive class prediction accuracy was used in the evaluation the effect of the prediction models. And a comprehensive index for the evaluation of the accuracy rate of models was obtained by combining the two indexes.(6) An integrated optimized model was used to improve prediction accuracy.(7) The proposed method was used in the empirical analysis for a political blog network and a purchase network, and compared with the traditional CN and RA methods, results showed that:the prediction accuracy of the proposed method is higher than the traditional methods, and the proposed method can be directly applied in bipartite networks.