NFE-PCN: A Node Feature Enhanced Embedding Framework for Pattern Change in Dynamic Network

Abstract

Dynamic networks are complex networks as their structures and node features change over time. However, they can better represent the real world, thus attracting the interest of researchers. Although realistic dynamic networks often exhibit changes in their patterns, the existing dynamic network models tend to classify all the snapshots as having the same pattern to learn during their embedding. These embedding models ignore a large amount of information about the patterns of dynamic networks. So, it is necessary to design a dedicated framework for learning the patterns of dynamic networks. Accordingly, this paper proposes a new framework, namely the NFE-PCN framework for effectively extracting information about the change in the patterns of networks. Specifically, the framework first determines the pattern in which the dynamic network snapshot is located, and then enhances the node information between networks by maintaining the same pattern. We conduct experiments with both real and artificial datasets for predicting links and classifying nodes. The obtained results show that the existing model under this framework decreases the computational effort in dynamic network embedding. The performance in the network embedding is improved by up to 29%, which is quite significant.

Keywords

• Dynamic network
• node feature
• snapshot
• link prediction
• graph neural network