Fast Heterogeneous Graph Neural Network Generation via Meta Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Heterogeneous Graph Neural Networks (HGNNs) are crucial for information extraction from complex graphs.
• Designing efficient HGNN architectures is challenging, requiring expertise and time for hyperparameter tuning.
• Existing Heterogeneous Graph Neural Architecture Search (HGNAS) methods lack generalizability for novel tasks.

Purpose of the Study:

• To develop an efficient and generalizable generative model for heterogeneous graph architectures.
• To overcome the limitations of task-specific HGNAS approaches.
• To enable rapid generation of optimal HGNN architectures for new, unseen tasks.

Main Methods:

• Introduced Heterogeneous Graph Neural Network Generation via Meta-Contrastive Learning (HGMCL).
• Utilized a meta-database of task-architecture pairs to learn a latent space.
• Employed meta-contrastive learning for efficient architecture generation based on task features.

Main Results:

• HGMCL successfully adapts to new tasks by generating optimal network architectures.
• The model significantly outperforms existing state-of-the-art methods in HGNN and HGNAS.
• Ablation studies confirmed the effectiveness of individual model components.

Conclusions:

• HGMCL offers an efficient and generalizable solution for automated HGNN architecture design.
• The meta-contrastive learning approach enables rapid adaptation to novel graph-related tasks.
• This work advances the field of automated machine learning for complex graph data.