Inductive Network Representation based on Neighbor Aggregation

Abstract:
Network representation learning is to learn low dimensional vectors for nodes. It plays a critical role in network analysis. However, most existing network embedding methods focus on embedding the nodes that already exist in the network, and cannot be generalized to newly-added nodes without retraining the learning model. To deal with this problem, the most common idea is to directly aggregate neighbor embeddings as the representation of new nodes, but this may lead to ignoring the particularity of node neighbors themselves. To fill this gap, we present INRL-NA, an innovative method of Inductive Network Representation Learning based on Neighbor Aggregation, which uses the hierarchical information to selectively aggregate neighbor node embeddings, preserving both the structural information and feature information of network. First we build a hierarchical network, then train a refinement model on the hierarchical network, finally after a new node joins the network, we use the hierarchical information and feature information to selectively aggregate neighbor embeddings. The experimental results show that compared with direct aggregation of all neighbor representations, INRL-NA has higher representation quality.