[Paper Review] GraphNAS: Graph Neural Architecture Search with Reinforcement Learning

Yang Gao, Hong Yang+3 more|arXiv (Cornell University)|Apr 22, 2019

Advanced Graph Neural Networks20 references70 citations

TL;DR

GraphNAS automatically searches for graph neural network architectures using a recurrent controller and reinforcement learning, achieving competitive or superior performance on node classification tasks compared to hand-designed models. It supports transductive and inductive settings and introduces parameter sharing to improve search efficiency.

ABSTRACT

Graph Neural Networks (GNNs) have been popularly used for analyzing non-Euclidean data such as social network data and biological data. Despite their success, the design of graph neural networks requires a lot of manual work and domain knowledge. In this paper, we propose a Graph Neural Architecture Search method (GraphNAS for short) that enables automatic search of the best graph neural architecture based on reinforcement learning. Specifically, GraphNAS first uses a recurrent network to generate variable-length strings that describe the architectures of graph neural networks, and then trains the recurrent network with reinforcement learning to maximize the expected accuracy of the generated architectures on a validation data set. Extensive experimental results on node classification tasks in both transductive and inductive learning settings demonstrate that GraphNAS can achieve consistently better performance on the Cora, Citeseer, Pubmed citation network, and protein-protein interaction network. On node classification tasks, GraphNAS can design a novel network architecture that rivals the best human-invented architecture in terms of test set accuracy.

Motivation & Objective

• Reduce manual effort in GNN design by automating architecture search.
• Leverage reinforcement learning to maximize validation accuracy of generated GNNs.
• Enable efficient search in a large GNN space with parameter sharing.
• Evaluate GraphNAS on both transductive and inductive node classification tasks.
• Demonstrate that GraphNAS designs can rival or surpass human-crafted architectures.

Proposed method

• Use a recurrent neural network as a controller to generate architecture descriptions as sequences of tokens.
• Define a flexible search space including sampling, correlation, aggregation, residual, and gating functions per layer.
• Train the controller with policy gradient to maximize expected validation accuracy of generated GNNs.
• Share parameters across generated child GNNs to accelerate training (with a careful sharing strategy that respects architectural compatibility).
• Train each sampled GNN on a validation set to obtain a reward signal for controller updates.
• Derive architectures by sampling multiple models from the trained controller and retraining the best candidate(s) from scratch.

Experimental results

Research questions

• RQ1 Can reinforcement learning effectively automate the search for graph neural architectures in a large, heterogeneous search space?
• RQ2 Does GraphNAS produce GNNs that outperform or rival hand-designed baselines on standard datasets in both transductive and inductive settings?
• RQ3 Does parameter sharing improve search efficiency without sacrificing final model quality?
• RQ4 What is the impact of different architectural choices (sampling, attention, aggregation, residual connections) on GNN performance?

Key findings

• GraphNAS consistently improves over several baselines on Cora, Citeseer, and PubMed in accuracy.
• On Cora, GraphNAS achieves 84.2% ±1.0% accuracy with a two-layer architecture.
• On Citeseer and PubMed, GraphNAS achieves 73.1% ±0.9% and 79.6% ±0.4% accuracy, respectively.
• In PPI inductive tasks, GraphNAS with no skip connections reaches 98.6% ±0.1% micro-F1, outperforming several strong baselines.
• Parameter sharing improves search efficiency and, in many cases, final architecture performance, vs. training from scratch.
• GraphNAS-attained architectures reach competitive or superior performance relative to human-invented models (e.g., GeniePath, GAT) on benchmark datasets.