Link prediction for multi-layer and heterogeneous cyber-physical networks

Cyber-physical networks with tight interactions between the cyber and physical components have gained wide popularity in recent years with the development of cyberspace. Network embeddings, which can generate relatively low-dimensional vectors, are commonly used to reveal the missing or potential patterns in complex cyber-physical networks. Due to the characteristics of heterogeneity and multiple layers, traditional embedding techniques, which focus on the co-occurrence of nodes in homogeneous and single layer networks, fail to represent the similarity between nodes far apart from each other and with distinct attributes. In this paper, a generalized framework termed Info2vec for graph embedding representation is proposed to predict the missing links in cyber-physical networks, where some links are not observed due to the incompletion of data in the mapping of cyberspace. This framework aggregates the identities of heterogeneous nodes in multiple layers from multiple perspectives, including local structure, spatial nearness, and nodal attributes, based on which a context graph is then generated. Numerous sequences of nodes are sampled from the context graph and used as the input of a language model, which will generate embedding vectors used for link prediction. We demonstrate the effectiveness of this approach in the prediction of missing links in a realistic cyber-physical information network, where it significantly outperforms many well-known baselines.