Abstract

The study of factors influencing the decision-making behavior of leaders is one of the key problems in social network analysis. In this paper, based on recurrent neural networks in deep learning to study the influence weights of different social network factors, an improvement method based on random perturbation of parameters is proposed for the problem of slow convergence of RNN and easy to fall into local optimal solutions, and random perturbation parameters are introduced in the hidden layer as a way to optimize the weight update function. Then, the sample data sets of network density, structural holes, centrality and linkage points are established and input to different deep learning networks for testing respectively. The classification accuracy of PRP-RNN is 7.38% and 6.45% higher than CNN and LSTM for network density, respectively. The classification accuracy rate for network structure hole is 9.67% and 5.35% higher for PRP-RNN than CNN and LSTM, respectively. The classification accuracy for network centrality is 10.30% and 7.36% higher for PRP-RNN than CNN and LSTM, respectively. The classification accuracy rate for network association points is 9.71% and 8.63% higher for PRP-RNN than CNN and LSTM, respectively. Based on the analysis of deep learning, the closed and dense social network is more conducive to narrowing the invisible minefield of decision making, and the leaders who occupy the structural hole and the central position of the network have easier access to diversified information and resources, which is more conducive to improving the quality of decision making.