Cooperative co-evolutionary search for meta multigraph and graph neural architecture on heterogeneous information networks

Abstract

To model the rich semantic information on heterogeneous information networks (HINs), heterogeneous graph neural architecture search (HGNAS) has become a research hotspot, as it offers a promising automatic search technique for heterogeneous graph neural networks (HGNNs). However, there is no method that can simultaneously solve the meta multigraph and neural architecture search, which are the two core problems of HGNAS. In addition, existing HGNAS methods can only search for the meta graph or determine the number of edge types by setting a threshold hyperparameter, which has limited expression or is difficult to determine and significantly affects performance. In this paper, a cooperative co-evolutionary meta multigraph and graph neural architecture search method (called CCMG) on HINs is proposed. Specifically, CCMG first represents the meta multigraph and neural architecture by discrete encodings, and the number of network layers is variable. Second, whether an encoding of the architecture is meaningful or not is affected by the value of the encoding taken at the corresponding meta multigraph position and their search space sizes are not imbalanced. To cope with these situations, they are cooperatively and collaboratively optimized in the form of subproblems, facilitating group collaboration and information sharing. Finally, the effectiveness and superiority of the CCMG are verified on six datasets for node classification and recommendation tasks. Over the comparison HGNAS method, CCMG improves its performance on the two tasks by an average of 2.29% and 1.21%, respectively.