HGNN⁺: General Hypergraph Neural Networks.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Existing Graph Neural Networks (GNNs) struggle with complex correlations in multi-modal/multi-type data due to reliance on simple graphs.
• Current hypergraph methods often concatenate individual modality graphs, hindering adaptive weighting and optimal fusion.
• There is a need for advanced frameworks capable of modeling high-order, multi-modal data correlations effectively.

Purpose of the Study:

• To introduce HGNN+, a generalized framework for high-order multi-modal/multi-type data correlation modeling.
• To develop a unified hypergraph-based approach that learns optimal data representations.
• To enable adaptive fusion of correlations from diverse data modalities within a single framework.

Main Methods:

• HGNN+ constructs hyperedge groups to capture latent high-order correlations within each modality.
• An adaptive hyperedge group fusion strategy is employed to integrate correlations from different modalities into a unified hypergraph.
• A novel spatial domain hypergraph convolution scheme is utilized for learning general data representations.

Main Results:

• HGNN+ consistently outperforms state-of-the-art methods across various datasets, particularly in modeling implicit data correlations.
• The framework demonstrates significant improvements in tasks such as data classification, retrieval, and recommendation.
• Evaluations confirm the efficacy of the adaptive fusion strategy and the hypergraph convolution scheme.

Conclusions:

• HGNN+ provides a powerful and flexible framework for advanced multi-modal data correlation modeling.
• The proposed approach effectively addresses the limitations of existing GNN and hypergraph-based methods.
• The release of the THU-DeepHypergraph toolbox facilitates practical applications of HGNN+.