Fast Heterogeneous Graph Neural Network Generation via Meta Contrastive Learning.

Abstract

Heterogeneous Graph Neural Networks (HGNNs) are extensively utilized for information extraction from intricate heterogeneous graphs. Despite their prevalence, crafting an efficient HGNN architecture is a formidable challenge, necessitating both extensive domain expertise and considerable expenditure of time to fine-tune network hyperparameters. Heterogeneous Graph Neural Architecture Search (HGNAS) algorithms offer a viable solution, automatically discovering optimal network architectures for predefined tasks within a given search space, and frequently surpassing the performance of manually designed models. However, conventional HGNAS approaches are constrained by their lack of generalizability, as they are tailored to specific tasks and do not easily adapt to novel ones. To counter this limitation, we introduce an efficient generative model for heterogeneous graph architecture, termed Heterogeneous Graph Neural Network Generation via Meta-Contrastive Learning (HGMCL). Our model leverages a meta-database comprising task-architecture pairs, which facilitates the learning of the task-architecture latent space. Confronted with a new task, HGMCL is capable of directly generating the optimal network architecture informed by task features in a few seconds. Empirical results indicate that our model not only successfully adapts to new tasks but also outperforms the majority of existing state-of-the-art methods in both heterogeneous graph neural networks and their architecture search. Furthermore, ablation studies substantiate the efficacy of each component within our model.