关键词: 不完整网络, 网络插补, 网络标签传播, 分类变量, 信用风险评估

Abstract: Traditional prediction methods usually model and predict the responses based on the covariate information, but seldom consider the network connection of individuals. However, the relationship of network nodes can provide information for the prediction of nodal responses. Based on this finding, this study proposes a network label propagation algorithm. Based on the framework of semi-supervised learning, this study takes the adjacency matrix as the basis of nodal similarity inference, then infers the response of unknown nodes through the connection information between nodes and the response of known nodes. The algorithm is suitable for incomplete data responses whose response variables are discrete. Under the assumption that the network follows the stochastic block model, this study proves that the response of unknown nodes can be predicted consistently by the algorithm. Numerical simulation and empirical research show that this algorithm performs well in prediction.

Key words: Incomplete Network, Network Imputation, Network Label Propagation, Categorical Variables, Credit Risk Rating