PaperlayerView exact revision on arXiv

Source-linked AI summary

SUCRe: Selective Uncertainty-Aware Contrastive Representation for Graph Transfer Learning

Mingcan Wang, Junchang Xin, Zhongming Yao, Bing Tian Dai, Kaifu Long, Zhiqiong Wang

arXiv:2609.30826v1cs.GR

TL;DR

Sparse target labels and graph shifts make reliable transfer challenging, while treating transferred samples uniformly leaves uncertainty and computational burden insufficiently addressed. SUCRe combines uncertainty-aware feature adaptation with selective semi-hard contrastive sampling. Across Table 1 settings, it ranks first or second in most cases, with a 1.32 average ranking versus the best baseline's 1.9.

  • Problem

    Graph-transfer methods often treat samples uniformly despite uncertainty from source-target graph differences, while transfer also carries computational burden.

    Source: Introduction · pp. 1–2

  • Method

    SUCRe uses entropy-based uncertainty-aware feature adaptation and domain-aware semi-hard negative selection to filter less informative or unreliable contrastive pairs.

    Source: Introduction · p. 2, Structure-Aware Entropy-Based Matching… · p. 3

  • Results

    Across Table 1 settings, SUCRe ranks first or second in most cases, with an average ranking of 1.32 versus the best baseline's 1.9.

    Source: Effectiveness. (RQ 1) · p. 5

  • Takeaways & Limitations

    The approach combines selective transfer with reduced unnecessary optimization, and experiments report effectiveness and efficiency for SUCRe and its modules.

    Source: Introduction · p. 2

Abstract

from arXiv · show

Graph transfer learning (GTL) provides a promising paradigm for adapting knowledge from source graphs with sufficient labels to label-scarce target graphs. However, existing approaches often assume that transferred knowledge is uniformly reliable, ignoring the different transferability of samples caused by structural and distribution shifts across graphs. This limitation leads to negative transfer and unnecessary computational overhead. In this work, we propose SUCRe, a selective uncertainty-aware contrastive representation method for GTL. The key idea is to selectively adapt and transfer graph knowledge according to its estimated reliability. Specifically, we introduce structure-aware entropy-based matching discrepancy, which jointly models feature uncertainty and structural coherence to ensure accurate feature adaptation between graphs. Moreover, we develop a domain-aware semi-hard negative sampling strategy that constructs informative contrastive sets by filtering unreliable cross-domain relationships, reducing computational redundancy while enhancing representation discrimination. Extensive experiments on graph transfer benchmarks demonstrate that SUCRe achieves competitive performance with improved efficiency.

Introduction

Label scarcity motivates graph transfer learning, but shifts between source and target graphs can undermine transfer. SUCRe addresses uncertainty and computational burden through uncertainty-aware feature adaptation and selective contrastive learning.

Introduction

  • Sparse labels limit supervised GNN representation learning, motivating transfer from labeled source graphs to label-scarce target graphs.GTL leverages shared structural patterns and feature representations to improve target learning under limited labels.Source: Introduction · p. 1
  • Source-target distribution shifts in node features and topology can cause negative transfer, sometimes making transfer worse than training from scratch.Source: Introduction · p. 1
  • Existing approaches treat transferred samples uniformly, leaving uncertainty under graph differences and the challenge of few-shot generalization insufficiently addressed.Source: Introduction · pp. 1–2
  • SUCRe pairs a small trainable GNN with a frozen pretrained GNN, using entropy-based feature adaptation and semi-hard negative selection.The approach targets effective transfer while avoiding samples that are too hard or too easy.Source: Introduction · p. 2
  • SEMD aligns graph information while considering uncertainty and preserving structural coherence; semi-hard sampling filters less informative or reliable samples to improve effectiveness and efficiency.The authors also report experiments on the overall framework and its modules.Source: Introduction · p. 2

Related Work

Related work spans graph-transfer and pretrained-model adaptation, while motivating methods that account for uncertainty and select useful contrastive negatives.

Graph Transfer Learning

  • Pretrain-fine-tune approaches adapt pretrained GNNs to downstream graphs with minimal supervision, and prior methods include GTOT, AdapterGNN, and GraphControl.The section also categorizes GTL as multi-task learning, multi-network learning, domain adaptation, and pretrain-fine-tune approaches.Source: Graph Transfer Learning · p. 2
  • MMD aligns transferable node embeddings but neglects structure, while SMMD incorporates structure but neglects node and feature uncertainty.SUCRe instead measures entropy of cross-domain similarity distributions to distinguish confident samples from ambiguous or noisy ones while preserving structural coherence.Source: Graph Transfer Learning · p. 2

Semi-hard Negative Sample Selection

  • Semi-hard negative sampling focuses on informative boundary cases while avoiding overly hard negatives, but its use in graph contrastive learning remains limited.SUCRe proposes domain-aware detection to filter less informative or unreliable negatives, whose use can waste resources or introduce noisy supervision under graph shifts.Source: Semi-hard Negative Sample Selection · p. 2

Methodology

SUCRe combines uncertainty-aware feature adaptation with selective contrastive learning to transfer graph knowledge while accounting for structural coherence and sample reliability. Its training objective adds classification and structure regularization, and its complexity analysis indicates greater efficiency than GraphLoRA.

Notation and Preliminaries

  • The notation uses G=(V,E) for a graph and X∈R^(f×N) for node features, with superscripts s and t distinguishing source and target.Each x_i is the feature vector for node v_i.Source: Notation and Preliminaries · p. 3

Framework Overview

  • An optional SVM-based feature-dimension reduction precedes SEMD, which adapts source and target features while preserving both graphs’ structural coherence.SEMD combines entropy-based discrepancy with source- and target-structure coherence terms.Source: Framework Overview · p. 3
  • SUCRe freezes the pretrained GNN and trains an additional GNN with contrastive learning, selecting semi-hard negatives to discard uninformative or unreliable pairs.Easy negatives offer limited discrimination, while extremely hard negatives may add ambiguity and noise.Source: Framework Overview · p. 3, Semi-Hard Negative Sample Detection… · p. 4
  • SEMD aligns source- and target-graph entropy distributions to model node-level transfer uncertainty under cross-graph shifts.Its structural weights preserve within-graph coherence; a PPR-based term prevents mapped features of similar target nodes from becoming too distant.Source: Structure-Aware Entropy-Based Matching… · pp. 3–4

Optimization

  • The optimization uses a node classifier and adds homophily-based structure regularization, with λ1, λ2, and λ3 weighting objective terms.The classification loss is defined for the classifier’s node predictions.Source: Optimization · p. 5

Complexity Analysis

  • SUCRe’s stated overall complexity is O(Nbf_t + NTf_t), compared with GraphLoRA’s O(KNbf_t + N^2f_t), and the paper says SUCRe tends to be more efficient.Here T<N and K>1; retaining T informative negatives per anchor reduces contrastive complexity from O(N^2d) to O(NTd).Source: Complexity Analysis · p. 5

Experiments

Experiments evaluate SUCRe across public, few-shot, source–target transfer, large-scale graph, efficiency, and ablation settings. SUCRe performs strongly against transfer baselines while retaining competitive memory and time costs.

Experimental Setting

  • Experiments use PubMed, Cora, CiteSeer, Amazon Photo, Amazon Computers, Physics, and WikiCS, comparing non-transfer and graph-transfer baselines.The baselines include GCN, PPNP, GPPT, DDSM, AdapterGNN, GraphControl, and GraphLoRA.Source: Datasets · p. 5, Baselines · p. 5

Effectiveness. (RQ 1)

  • On WikiCS and Physics, SUCRe outperforms GraphLoRA and GraphControl in 10-, 20-, and 30-shot settings.The t-SNE comparison on CiteSeer with a Cora-pretrained model reports a Silhouette score 1.15 times GraphLoRA’s and a Davies–Bouldin index 0.805 times GraphLoRA’s.Source: Effectiveness. (RQ 1) · pp. 5–6, Figure 2 · p. 6, Figure 5 · p. 7
  • SUCRe improves over GraphLoRA by up to 3.4% in public settings and 4.18% in few-shot settings, with average rank 1.32 versus the best baseline’s 1.9.Table 1 covers public, few-shot, and different source-target node-classification settings.Source: Effectiveness. (RQ 1) · p. 5, Table 1 · p. 6

Efficiency. (RQ 2)

  • SUCRe uses substantially less GPU memory than GraphLoRA, with memory use nearly unchanged as feature-reduction output channels increase.It also requires less time than GraphLoRA under the same settings, though it is not always the fastest method; GraphControl’s lower time and memory come with substantially inferior accuracy.Source: Efficiency. (RQ 2) · p. 6, Figure 3 · p. 7, Figure 4 · p. 7

Ablation Study. (RQ 3)

  • SUCRe outperforms its ablation variants in most cases, supporting contributions from feature adaptation, SEMD, and semi-hard negative mining.The variants remove feature adaptation, replace SEMD with MMD or SMMD, or remove semi-hard mining.Source: Ablation Study. (RQ 3) · p. 7, Table 2 · p. 7

Conclusion

SUCRe combines uncertainty-aware feature adaptation with selective contrastive learning for graph transfer. It achieves competitive or superior benchmark performance, especially in few-shot scenarios, while using less computation and memory.

Conclusion

  • SUCRe consistently achieves competitive or superior performance to existing graph transfer methods, especially in few-shot scenarios.The conclusion also reports lower computational and memory costs.Source: Conclusion · p. 7
  • Uncertainty-aware feature adaptation adaptively matches feature distributions between source and target graphs.Source: Conclusion · p. 7
  • Semi-hard mining selects informative samples to improve transfer effectiveness while reducing unnecessary optimization.Source: Conclusion · p. 7

Related papers

Loading 2609.30826v1…