Title:Robust and Efficient Network Reconstruction in Complex System via Adaptive Signal Lasso

Abstract:Network reconstruction is important to the understanding and control of collective dynamics in complex systems. Most real networks exhibit sparsely connected properties, and the connection parameter is a signal (0 or 1). Well-known shrinkage methods such as lasso or compressed sensing (CS) to recover structures of complex networks cannot suitably reveal such a property; therefore, the signal lasso method was proposed recently to solve the network reconstruction problem and was found to outperform lasso and CS methods. However, signal lasso suffers the problem that the estimated coefficients that fall between 0 and 1 cannot be successfully selected to the correct class. We propose a new method, adaptive signal lasso, to estimate the signal parameter and uncover the topology of complex networks with a small number of observations. The proposed method has three advantages: (1) It can effectively uncover the network topology with high accuracy and is capable of completely shrinking the signal parameter to either 0 or 1, which eliminates the unclassified portion in network reconstruction; (2) The method performs well in scenarios of both sparse and dense signals and is robust to noise contamination; (3) The method only needs to select one tuning parameter versus two in signal lasso, which greatly reduces the computational cost and is easy to apply. The theoretical properties of this method are studied, and numerical simulations from linear regression, evolutionary games, and Kuramoto models are explored. The method is illustrated with real-world examples from a human behavioral experiment and a world trade web.