Topology Identification for Complex Dynamical Networks Based on XGBoost Using Multiple Feature Sequences

Efficient method of topology identification is the basis for studying and controlling complex dynamical networks. In this paper, a novel data-driven identification method called MFE-XGBoost based on multiple feature extractors and extreme gradient boosting (XGBoost) is proposed. We use multiple feature extractors to extract latent features of different dimensions from time series of node states, then construct classification and regression tree (CART) in XGBoost using feature sequences. We binarize the information gain matrix calculated by CARTs, then obtain the adjacency matrix of the network to realize topology identification. The results show that, compared with other methods, MFE-XGBoost has higher accuracy and better anti-interference performance, because it considers the characteristics of more dimensions and designs a reasonable threshold selection method for binarization. MFE-XGBoost requires shorter time series to achieve a good identification effect because of the good fitting ability of XGBoost.