AutoMaster: Differentiable Graph Neural Network Architecture Search for Collaborative Filtering Recommendation

Caihong Mu, Haikun Yu, Keyang Zhang, Qiang Tian, Yi Liu

Abstract: Graph Neural Networks (GNNs) have been widely applied in Collaborative Filtering (CF) and have demonstrated powerful capabilities in recommender systems (RSs). In recent years, there has been a heated debate on whether the non-linear propagation mechanism in Graph Convolutional Networks (GCNs) is suitable for CF tasks, and the performance of linear propagation is believed to be superior to non-linear propagation mainly in the field of RSs. Therefore, it is necessary to reexamine this issue: (1) whether linear propagation generally outperforms non-linear propagation, and (2) whether a combination of linear and non-linear propagation can be applied to CF tasks to achieve better accuracy. Furthermore, most existing studies design a single model architecture tailored to specific data or scenarios, and there remains a challenging and worthwhile problem to obtain the best-performing model in new recommendation data. To address the above issues, we propose a model called AutoMaster, which implements differentiable graph neural network architecture search for CF recommendation and automatically designs GNN architectures specific to different datasets. We design a compact and representative search space that includes various linear and non-linear graph convolutional layers, and employ a differentiable search strategy to search for the best-performing hybrid architecture in different recommendation datasets. Experimental results on five real-world datasets demonstrate that the GNN automatically achieved by the proposed AutoMaster contains both linear and nonlinear propagation, and outperforms several advanced GNN based CF models designed by the experienced human designers.