Fine-Tuned Heterogeneous Graph Convolutional Network Embedding

Abstract

Heterogeneous network representation learning, which can understand and capture the complex interactions between different nodes and edges in heterogeneous networks, is a research field that has garnered widespread attention. However, existing models often overlook the varying effects of meta-paths on target nodes, even when they share the same relationship type. This limitation makes it difficult to accurately represent heterogeneous structures. To solve this issue, we introduce FHGCN, a novel Fine-tuned Heterogeneous Graph Convolutional Network that leverages structural information and attribute semantics to enrich the learned node embedding and improve the final performance. The model consists of two parts: A fine-tuned multi-relation aggregation module which automatically learns the weights of heterogeneous relations and further deduces the significance of meta-paths by considering the distinct fluctuation in the impact of meta-paths generated by the same relation. And a multilayer convolutional module which can learn valuable interactions between heterogeneous meta-paths of different lengths to obtain favorable node representation node embedding. Node classification experiments on two heterogeneous graph datasets show that FHGCN has a good predictive ability.