How GraphKeeper Tackles Catastrophic Forgetting in Domain‑Incremental Graph Learning

Problem Overview

Catastrophic forgetting is the core issue in incremental learning: when a model learns new tasks or data, it tends to forget previously acquired knowledge. Graph incremental learning (GIL) aims to continuously integrate new graph knowledge without retraining the entire network, while preserving recognition of old graph data.

Domain‑IL Challenge

Existing GIL methods focus on Task‑IL and Class‑IL settings, where new tasks or classes are confined to a single graph domain. In realistic scenarios, incremental graphs may come from different domains, making domain‑incremental learning (Domain‑IL) essential. Current GIL approaches are designed for single‑domain cases and cannot handle the unique challenges of Domain‑IL.

Analysis of Forgetting

Two factors cause forgetting in graph models when learning new graphs: (1) shift of previously learned graph embeddings, and (2) bias on the decision boundary. Parameter changes during adaptation lead to embedding drift and decision‑boundary deviation, especially across domains with large structural and semantic gaps.

Proposed Method: GraphKeeper

GraphKeeper addresses the two identified problems with three key components:

Multi‑domain graph decoupling : isolates parameters of each domain using a domain‑specific LoRA module attached to a frozen pre‑trained GNN.

Unbiased knowledge preservation : fits class labels on stable embeddings via ridge regression, avoiding gradient updates that shift the decision boundary.

Domain‑aware distribution discrimination : matches test graphs to domain prototypes to infer the correct domain when it is unobservable.

Multi‑Domain Graph Decoupling

A GNN is pre‑trained once. For each incremental domain i, a low‑rank LoRA module \(\mathbf{A}_i, \mathbf{B}_i\) (rank \(r \ll d\)) is added. The i‑th layer output is:

During learning of a new domain, LoRA parameters of previous domains are frozen, keeping their embeddings stable.

In‑Domain Decoupling Objective

Within a single domain, a contrastive loss encourages nodes of the same class to be close and nodes of different classes to be far:

Cross‑Domain Decoupling Objective

For each previously seen domain, compact embedding prototypes are obtained by clustering. The current domain is pushed away from all prior prototypes:

Unbiased Knowledge Preservation

Instead of jointly updating the classifier with the embedding model, GraphKeeper freezes the embedding and learns a ridge‑regression classifier on the stable embeddings. For the first domain, the optimal weight \(\mathbf{W}\) satisfies:

For subsequent domains, a recursive update yields the optimal parameters without accessing historical data:

Domain‑Aware Distribution Discrimination

Test graphs are first mapped to a high‑dimensional space using a frozen random GNN, producing domain prototypes. The test graph is assigned to the nearest prototype based on Euclidean distance, enabling accurate inference for unseen domains.

Experiments

Eight incremental groups were constructed from 15 real‑world datasets, covering homogeneous, mixed, and long‑sequence domain scenarios. GraphKeeper consistently outperformed baselines in Domain‑IL, improving average accuracy by 6.5%–16.6% and reducing average forgetting to near zero. In longer incremental sequences, it still led by 22.8%–23.7%.

Ablation Study

Removing any component degrades performance:

Without the decoupling targets (w/o DT), embeddings from different domains and classes become confused.

Without domain‑specific LoRA (w/o PEFT), new‑domain adaptation shifts embeddings and harms knowledge preservation.

Without unbiased knowledge preservation (w/o KP), the classifier updates jointly with embeddings, causing decision‑boundary bias.

Embedding Visualization

GraphKeeper produces compact, well‑separated embeddings across domains, while baselines such as PDGNN and DeLoMe show significant overlap and increasing confusion over time.

Conclusion

GraphKeeper introduces a novel GIL framework that mitigates embedding shift and decision‑boundary bias, effectively solving catastrophic forgetting in domain‑incremental graph learning. Limitations include the assumption of completely new classes per domain and the lack of exploitation of zero‑shot capabilities; future work will address more general cross‑domain scenarios.