Automatic search of architecture and hyperparameters of graph convolutional networks for node classification

Graph neural networks (GNNs) rely heavily on architecture design and artificial hyperparameters, often resulting in expensive manual effort and poor performance. Recently, automated machine learning (AutoML) on graphs is a promising approach for automated neural network design that is gaining attention from the research community. The existing AutoML methods for GNNs usually preset the number of layers to a small, fixed number, e.g., no more than three layers, possibly due to the oversmoothing problem. However, it is well known that deeper network models facilitate the extraction of high-level information representations. To this end, using the genetic algorithm (GA) as a framework, in this paper we propose a GNN search method (called GCN-GA) that dynamically searches the depth of the model to efficiently handle the node classification tasks. First, a variable-length encoding strategy is proposed to use the constituent units of four graph convolutional network (GCN) structures with different topologies as building blocks for representing the network architecture. Second, considering that the search method of searching network architectures with fixed hyperparameters and then searching hyperparameters independently may result in the obtained network models be not optimal, GCN-GA represents hyperparameters in the form of fixed-length encoding. They form a hybrid encoding strategy for representing the network architecture and hyperparameters. Then, the GCN-GA simultaneously searches both the network architecture and hyperparameters during the evolution. In addition, for the variable-length encoding, an improved two-point crossover operator, and three types of variation operators are designed for the evolutionary process. Finally, experiments are conducted on three widely used node classification datasets, namely, Cora, CiteSeer, and PubMed, in the semi-supervised and supervised tasks. The experimental results show that GCN-GA achieves more effective classification accuracy in most cases compared with the state-of-the-art hand-designed GNNs and the AutoML methods. The code will be available at https://github.com/chnyliu/GCN-GA .