Pre-training Graph Neural Networks with Kernels

  • 2018-11-16 17:24:58
  • Nicolò Navarin, Dinh V. Tran, Alessandro Sperduti
  • 1

Abstract

Many machine learning techniques have been proposed in the last few years toprocess data represented in graph-structured form. Graphs can be used to modelseveral scenarios, from molecules and materials to RNA secondary structures.Several kernel functions have been defined on graphs that coupled withkernelized learning algorithms, have shown state-of-the-art performances onmany tasks. Recently, several definitions of Neural Networks for Graph (GNNs)have been proposed, but their accuracy is not yet satisfying. In this paper, wepropose a task-independent pre-training methodology that allows a GNN to learnthe representation induced by state-of-the-art graph kernels. Then, thesupervised learning phase will fine-tune this representation for the task athand. The proposed technique is agnostic on the adopted GNN architecture andkernel function, and shows consistent improvements in the predictiveperformance of GNNs in our preliminary experimental results.

 

Quick Read (beta)

loading the full paper ...