Network Structure and Feature Learning from Rich but Noisy Data.

In the study of network structures, much attention has been devoted to network reconstruction, which relies on partial edge-related information or dynamical processes on the network. However, there are cases where we are only given incomplete nodal data, and the nodal data are measured with different methodologies. In this work, we present an unsupervised learning framework to construct networks from noisy and heterogeneous nodal data. First, we introduce the creating nodes' context sets, which are used to generate random node sequences. Then, a three-layer neural network is adopted to train the node sequences to infer node vectors, enabling us to capture nodes with synergistic roles within the network. Further, the effectiveness of the method is validated through both synthetic data and real data. Finally, we compare the differences between the global thresholding method and the entropy-based method in edge selection. In summary, this work presents a neural network method for node vector learning from heterogeneous nodal data and an entropy-based method for edge selection.