Training-free Graph Neural Networks and the Power of Labels as Features

TL;DR

This paper introduces training-free graph neural networks that leverage labels as features to enhance expressiveness and achieve superior performance with minimal training, offering a new paradigm for node classification.

Contribution

It proposes training-free GNNs utilizing labels as features, demonstrating improved expressiveness and efficiency over traditional GNNs.

Findings

TFGNNs outperform existing GNNs in training-free settings.

TFGNNs converge faster with fewer training iterations.

Label as features (LaF) enhances GNN expressiveness.

Abstract

We propose training-free graph neural networks (TFGNNs), which can be used without training and can also be improved with optional training, for transductive node classification. We first advocate labels as features (LaF), which is an admissible but not explored technique. We show that LaF provably enhances the expressive power of graph neural networks. We design TFGNNs based on this analysis. In the experiments, we confirm that TFGNNs outperform existing GNNs in the training-free setting and converge with much fewer training iterations than traditional GNNs.