Abstract
The network alignment task aims to identify the correspondence of the same nodes across networks, which plays a crucial role in many fields. However, most existing methods focus solely on consistency assumptions at the local structure information of nodes, meaning that the alignment of neighboring node pairs tends to be consistent with each other, while ignoring information from other scales, such as higher-order information. This can easily lead to over-smoothing issues. In this paper, we propose a novel multi-scale embedding-based network alignment method in this paper. Specifically, we capture three types of node embedding information, including first-order embedding, multi-order embedding, and higher-order embedding, to obtain rich information. Then, we enhance the multi-order and higher-order embedding information of nodes by applying a stable-pair based fine-tuning method. Finally, we use a fusion mechanism based on importance weights to combine the alignment matrices obtained from these three types of node embedding information and obtain the optimal alignment matrix. We conduct extensive experiments on three real-world datasets to verify the effectiveness of our proposed model.
