Link Prediction in Dynamic Heterogeneous Networks via a Novel Embedding Technique

Abstract

Link prediction via network embedding techniques has been verified to be efficient and reliable in real applications. However, most current studies focus on the link prediction in static homogeneous networks. We aim to predict dynamic links in heterogeneous networks. Most network embedding techniques only have preserved the feature of low or high-order proximity with neighbors. In this paper, we propose a novel embedding technique that combines the proximity feature of node’s neighbors as well as the structural feature of node’s community. Our method is consisted by two levels. The first-level learning is a dynamic embedding algorithm, which enables a node to learn the proximity feature from its neighbors in a continuous snapshot of the network. The second-level makes the node to learn the structural feature from the nodes in the same community. We test our method on a real-world dataset which contains the customer-commodity relations in continuous network snapshots. Experimental result shows that our method outperforms the original method and other state-of-art methods baselines.