Type Preserving Representation of Heterogeneous Information Networks

Abstract

Listen

Translate article

In the current information explosion era, many complex systems can be modeled using networks/graphs. The development of artificial intelligence and machine learning has also provided more means for graph analysis tasks. However, the high-dimensional large-scale graphs cannot be used as input to machine learning algorithms directly. One typically needs to apply representation learning to transform the high-dimensional graphs to low-dimensional vector representations. As for network embedding/representation learning, the study on homogeneous graphs is already highly adequate. However, heterogeneous information networks are more common in real-world applications. Applying homogeneous-graph embedding methods to heterogeneous graphs will incur significant information loss. In this paper, we propose a numerical signature based method, which is highly pluggable—given a target heterogeneous graph G, our method can complement any existing network embedding method on either homogeneous or heterogeneous graphs and universally improve the embedding quality of G, while only introducing minimum overhead. We use real datasets from four different domains, and compare with a representative homogeneous network embedding method, a representative heterogeneous network embedding method, and a state-of-the-art heterogeneous network embedding method, to illustrate the improvement effect of the proposed framework on the quality of network embedding, in terms of node classification, node clustering, and edge classification tasks.