In this paper, we proposed to tackl

In this paper, we proposed to tackle the problem of attributed network embedding, which involved learning low-dimensional representations of nodes that preserve both the structure and the attribute information. We proposed a DANE framework that tackled the data sparsity, structure and attribute preserving, and nonlinearly patterns of attributed network embedding in a unified framework. Specifically, the DANE framework is composed of three steps. First, a step-based random walk is proposed to capture the interaction between network structure and node attributes from various degrees of proximity. Then, we constructed an enhanced matrix representation of the attributed network by summarizing the various degrees of proximity. In the third step, we designed a deep neural network to exploit the complex, and nonlinear patterns in the enhanced matrix for network embedding. We conducted extensive experimental results on various datasets, and the results clearly showed the superiority of our proposed DANE framework compared to the state-of-the-art baselines.In the future, we would like to explore and extend the proposed DANE framework for attributed network embedding models from the following two directions. First, we would improve the efficiency of DANE by learning to hash techniques, such that it could be applied to large-scale industrial scenarios. Second, as the network structure evolves over time, new edges come and old edges disappear. The incoming nodes may be incomplete with missing links or missing attributes. We plan to design the incremental algorithms for attributed network embedding as a future direction.