Learning Graph Structure for GNNs via Marginal Likelihood

Anita Yang, Thomas Möllenhoff, Ken-Ichi Kawarabayashi, Mohammad Emtiyaz Khan

Keywords: Graph Structure Learning, Graph Neural Networks, Marginal Likelihood, Laplace Approximation

Abstract: Learning graph structures for Graph Neural Networks (GNNs) can improve their performance, but it is challenging to design a good structure-learning objective that can provide sufficient flexibility while avoiding overfitting. Here, we propose to use marginal likelihood to learn the graph structure. We adopt the standard Laplace's method to approximate the marginal likelihood and optimize it using the Straight Through Estimator. This simple scheme yields good results on standard benchmarks for graph structure learning. The method can learn generic graph structures and eliminates the effort needed to design the objective. Our work makes it easier to learn graph structures for GNNs.

Submission Number: 12