Network embedding based on deep extreme learning machine

Network embedding, which learns low-dimensional representations for each node with the goal of capturing and preserving the complex structure of original networks, has shown its necessity in network analysis. The structure of real-world networks is highly non-linear; however, most existing methods cannot be well applied due to their shallow models. While a few deep neural networks have been adopted to capture the highly non-linearity, the deep structure makes them difficult to optimize in practice. In this paper, we propose a novel deep network embedding method, which exploits the fast learning speeds of extreme learning machine (ELM). Particularly, we first design a deep ELM-based auto-encoder, based on which we then proposed an extended model to preserve both first-order and second-order proximities by a joint loss function. Extensive experiments on real-world network datasets show the effectiveness and efficiency of proposed method as compared to state-of-the-art embedding methods by network recovery, multi-class classification and multi-label classification tasks.