HeterSEED: New Framework for Heterogeneous Graph Learning Under Heterophily
A novel framework named HeterSEED tackles the issue of heterophily in the context of heterogeneous graph learning. Heterophily arises when connected nodes possess differing labels or semantic roles, leading traditional graph neural networks to disseminate inaccurate information. HeterSEED separates representation learning into two distinct channels: a heterogeneous semantic channel that captures local semantics informed by type and relation, and a structure-aware heterophily channel that differentiates between homophilic and heterophilic neighborhoods through pseudo-label-guided partitioning, utilizing metapath-based structural weights for aggregation. Subsequently, an adaptive fusion mechanism at the node level merges the results. This research is available on arXiv with the identifier 2605.04594.
Key facts
- HeterSEED is a semantics-structure decoupling framework for heterogeneous graph learning under heterophily.
- It addresses heterophily, where connected nodes have dissimilar labels or semantic roles.
- Standard heterogeneous graph neural networks can propagate misleading information under heterophily.
- HeterSEED has two channels: a heterogeneous semantic channel and a structure-aware heterophily channel.
- The heterophily channel uses pseudo-label-guided partitioning to separate homophilic and heterophilic neighborhoods.
- Aggregation uses metapath-based structural weights.
- A node-level adaptive fusion mechanism combines outputs from both channels.
- The paper is available on arXiv with ID 2605.04594.
Entities
Institutions
- arXiv