Mitigating Negative Transfer in Multi-Domain Graph Learning

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 140 papers52 conceptsfrozen 2026-10-04

In the era of Large Language Models, the "scaling law" suggests that more data leads to better generalization. However, in the world of Graph Neural Networks (GNNs), researchers have encountered a stubborn obstacle: negative transfer. Unlike text, where a "universal grammar" exists, graphs from different domains—such as social networks, molecular structures, and citation maps—often possess contradictory structural logics. When a model is trained on these diverse sources simultaneously, the knowledge from one domain can actively degrade performance on another.

The Scaling Paradox and the Curse of Big Data

In traditional deep learning, pre-training on a massive, diverse corpus is the "holy grail." But as Jiarong Xu and Yuxuan Cao identify in Better with Less: A Data-Active Perspective on Pre-Training Graph Neural Networks, graph learning suffers from a "curse of big data." Adding more unlabeled data often introduces structural noise that is incompatible with the target task, leading to worse performance than training from scratch.

This phenomenon occurs because GNNs are hypersensitive to topology. As Zehong Wang and Chuxu Zhang explain in Tackling Negative Transfer on Graphs, minor structural differences between graphs can wildly distort node embeddings. When a model optimized for the homophily of a social network (where similar nodes connect) is applied to a biological network (where dissimilar nodes might connect), the transferred inductive biases become liabilities.

Why Graphs Resist Generalization

The core of the problem lies in the lack of a universal graph vocabulary. In Graph Foundation Models Are Already Here, Haitao Mao and Yao Ma argue that while language has words, graphs lack basic transferable units that encode structural invariance across domains.

Without this vocabulary, models struggle with:

  1. Structural Misalignment: The "relational logic" of e-commerce graphs does not align with the physical constraints of molecules.
  2. Feature Heterogeneity: Raw feature dimensions often represent entirely different things across datasets, leading models to align noise rather than meaning Yujing Liu.
  3. Catastrophic Forgetting: As shown by Antonio Carta in Catastrophic Forgetting in Deep Graph Networks, GNNs tend to overwrite structural patterns learned from initial tasks when exposed to new domains with different topological properties.

Strategies for Synergistic Learning

To turn domain conflict into synergy, recent research focuses on three primary mechanisms:

1. Discretizing the Graph Vocabulary

Instead of treating graphs as continuous matrices, researchers are moving toward "tokenizing" graph structures. Zehong Wang introduced GFT, which treats local message-passing structures (computation trees) as discrete tokens. By quantizing these trees into a transferable vocabulary, the model can recognize shared structural motifs across vastly different domains, such as citation networks and molecular graphs.

2. Domain-Aware Alignment

Rather than forcing all graphs into a single rigid representation, frameworks like MDGPT by Xingtong Yu and Chang Zhou use learnable domain tokens to align disparate semantic spaces. Similarly, Haihong Zhao proposed GCOPE, which uses virtual "coordinator" nodes to bridge different datasets. These methods allow the model to distinguish between domain-specific nuances and general graph features, preventing the "blurring" of representations that causes negative transfer.

3. Relational Fingerprinting

To bypass the noise of heterogeneous raw features, Yujing Liu and Shirui Pan developed REFI-GAD. This approach replaces raw dimensions with a universal "Relational Fingerprint"—a low-dimensional structural signature. By focusing on how nodes relate to their neighbors rather than what the nodes' attributes are, the model achieves a level of feature-agnosticism that is crucial for cross-domain transfer.

The Path to Graph Foundation Models

The transition from specialized GNNs to Graph Foundation Models (GFMs) depends on solving the negative transfer paradox. As Jiajun Zhu and Pan Li suggest in When Do Graph Foundation Models Transfer?, the goal is not just to aggregate more data, but to develop architectures that can reconcile conflicting structural information. By moving toward discrete vocabularies and domain-aware adapters like GraphLoRA, the field is beginning to unlock the same scaling potential seen in other areas of AI.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mitigating_negative_transfer_in_multi_do_1a5b2b86,
  title = {Mitigating Negative Transfer in Multi-Domain Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-negative-transfer-in-multi-domain-graph-learning/rdr_1a5b2b86b9d64646abb3b100e818c1e2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_mitigating_negative_transfer_in_multi_do_1a5b2b86_v1,
  title = {Mitigating Negative Transfer in Multi-Domain Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-negative-transfer-in-multi-domain-graph-learning/rdr_1a5b2b86b9d64646abb3b100e818c1e2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}