Automated Graph Representation Learning for Node Classification

Graphs are ubiquitous and play an essential role in the real world. A vital prerequisite for analyzing graphs is to learn their effective representations. Most existing graph representation learning models are hand-crafted, which lack the scalability to different kinds of graphs. In this paper, we present AutoGRL, an AUTOmated Graph Representation Learning Framework for node classification. We first design an appropriate search space with four critical components in the automated machine learning (AutoML) pipeline: data augmentation, feature engineering, hyper-parameter optimization, and neural architecture search. Then we search for the best graph representation learning model in the search space on given graph data using an efficient searching algorithm. We conduct extensive experiments on four real-world node classification datasets to demonstrate that AutoGRL can automatically find competitive graph representation learning models on specific graph data effectively and efficiently.