DRSA: Decoupled Relation Subspace Alignment for Heterogeneous Graph Foundation Models

Paper: Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment Authors: Ziyu Zheng, Yaming Yang, Zhe Wang, Ziyu Guan, Wei Zhao arXiv: 2605.00731 | 2026-04-30

The Problem: One-Size-Fits-All Graph Models

Traditional graph foundation models force all nodes - users, posts, tags, locations in a social network, for example - into a single shared feature space. User interests and location coordinates end up in the same dimensions; the semantics of "follow" and "like" relations get mixed together. The result is type collapse and relation confusion: type-specific semantics are distorted and original topology is destroyed.

Two Core Challenges for Heterogeneous Graph Foundation Models

Heterogeneous graphs contain multiple types of nodes and edges. Scaling them into foundation models faces:

1. Cross-Type Feature Shifts - different node types have different feature distributions; forcing alignment distorts each type's unique semantics. 2. Intra-Domain Relation Gaps - even within the same domain, different relations carry different semantics; global alignment ignores relation-level nuances.

Existing methods (global feature alignment via PCA or SVD) blindly force a shared feature space, leading to semantic distortion.

DRSA: Decoupled Relation Subspace Alignment

Core idea: Don't align all types into one space. Preserve an independent subspace for each relation type, and handle alignment in a decoupled manner.

Key design:

1. Relation decoupling - identify different relation types and learn an independent subspace for each, avoiding one-size-fits-all alignment. 2. Subspace alignment - align within each subspace, preserving type-specific semantics while achieving cross-domain transfer. 3. Plug-and-play - DRSA is a generic module that can be inserted into any graph foundation model without redesigning the architecture.

Analogy: organize a multilingual conference not by forcing everyone into a single mixed language, but by keeping an independent translation channel for each language with precise translation when needed.

Why It Matters

Multi-domain heterogeneous graphs (MDHGs) are ubiquitous:

• Social networks: users-posts-tags-locations
• Knowledge graphs: entities-relations-attributes
• Recommendation systems: users-items-categories-brands
• Biomedicine: genes-proteins-drugs-diseases
• Academic networks: authors-papers-institutions-topics

In all these scenarios, node types have fundamentally different semantics and relation types convey different information - blanket alignment loses key information. DRSA lets graph foundation models treat heterogeneity as a property to respect rather than a nuisance.

Key Insight

The global-alignment assumption is that a universal semantic space exists into which all node types can be mapped. DRSA's insight: perhaps no such universal space exists. Better to understand each type in its own subspace, then coordinate at a higher level.

Questions worth asking when handling multi-modal or multi-type data:

1. Does my alignment method respect the uniqueness of different types/modalities? 2. Is one-size-fits-all alignment losing critical information? 3. Could decoupled alignment improve representation quality? 4. Is a plug-and-play modular design applicable to my scenario?

In representation learning, "unification" is not always good. Sometimes "separate, then coordinate" beats "forced unification" - real understanding begins with acknowledging differences.