Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DREAM: Dual-Standard Semantic Homogeneity with Dynamic Optimization for Graph Learning with Label Noise

Yusheng Zhao, Jiaye Xie, Qixin Zhang, Weizhi Zhang, Xiao Luo, Zhiping Xiao, Philip S. Yu, Ming Zhang

TL;DR

DREAM tackles graph learning with label noise by introducing a dual-standard semantic homogeneity framework that uses proximity- and topology-informed anchors to measure node reliability. The semantic homogeneity scores reweight the supervision in an iterative, dynamic optimization, enabling robust, relation-aware training. The authors provide a theoretical guarantee for an epsilon-approximation under standard optimization assumptions and demonstrate strong empirical performance across six datasets and multiple noise types, outperforming both Euclidean-data LLN methods and topology-agnostic baselines. This topology-aware, anchor-based reliability reweighting offers a practical and scalable approach to robust graph learning in noisy real-world settings.

Abstract

Graph neural networks (GNNs) have been widely used in various graph machine learning scenarios. Existing literature primarily assumes well-annotated training graphs, while the reliability of labels is not guaranteed in real-world scenarios. Recently, efforts have been made to address the problem of graph learning with label noise. However, existing methods often (i) struggle to distinguish between reliable and unreliable nodes, and (ii) overlook the relational information embedded in the graph topology. To tackle this problem, this paper proposes a novel method, Dual-Standard Semantic Homogeneity with Dynamic Optimization (DREAM), for reliable, relation-informed optimization on graphs with label noise. Specifically, we design a relation-informed dynamic optimization framework that iteratively reevaluates the reliability of each labeled node in the graph during the optimization process according to the relation of the target node and other nodes. To measure this relation comprehensively, we propose a dual-standard selection strategy that selects a set of anchor nodes based on both node proximity and graph topology. Subsequently, we compute the semantic homogeneity between the target node and the anchor nodes, which serves as guidance for optimization. We also provide a rigorous theoretical analysis to justify the design of DREAM. Extensive experiments are performed on six graph datasets across various domains under three types of graph label noise against competing baselines, and the results demonstrate the effectiveness of the proposed DREAM.

DREAM: Dual-Standard Semantic Homogeneity with Dynamic Optimization for Graph Learning with Label Noise

TL;DR

DREAM tackles graph learning with label noise by introducing a dual-standard semantic homogeneity framework that uses proximity- and topology-informed anchors to measure node reliability. The semantic homogeneity scores reweight the supervision in an iterative, dynamic optimization, enabling robust, relation-aware training. The authors provide a theoretical guarantee for an epsilon-approximation under standard optimization assumptions and demonstrate strong empirical performance across six datasets and multiple noise types, outperforming both Euclidean-data LLN methods and topology-agnostic baselines. This topology-aware, anchor-based reliability reweighting offers a practical and scalable approach to robust graph learning in noisy real-world settings.

Abstract

Graph neural networks (GNNs) have been widely used in various graph machine learning scenarios. Existing literature primarily assumes well-annotated training graphs, while the reliability of labels is not guaranteed in real-world scenarios. Recently, efforts have been made to address the problem of graph learning with label noise. However, existing methods often (i) struggle to distinguish between reliable and unreliable nodes, and (ii) overlook the relational information embedded in the graph topology. To tackle this problem, this paper proposes a novel method, Dual-Standard Semantic Homogeneity with Dynamic Optimization (DREAM), for reliable, relation-informed optimization on graphs with label noise. Specifically, we design a relation-informed dynamic optimization framework that iteratively reevaluates the reliability of each labeled node in the graph during the optimization process according to the relation of the target node and other nodes. To measure this relation comprehensively, we propose a dual-standard selection strategy that selects a set of anchor nodes based on both node proximity and graph topology. Subsequently, we compute the semantic homogeneity between the target node and the anchor nodes, which serves as guidance for optimization. We also provide a rigorous theoretical analysis to justify the design of DREAM. Extensive experiments are performed on six graph datasets across various domains under three types of graph label noise against competing baselines, and the results demonstrate the effectiveness of the proposed DREAM.
Paper Structure (18 sections, 1 theorem, 12 equations, 6 figures, 3 tables, 1 algorithm)

This paper contains 18 sections, 1 theorem, 12 equations, 6 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Graph Neural Networks
  4. Learning with Label Noise
  5. Methodology
  6. Problem Definition
  7. Framework Overview
  8. Proximity-Aware Anchor Selection
  9. Topology-Aware Anchor Selection
  10. Relation-Informed Dynamic Optimization
  11. Theoretical Analysis
  12. Experiments
  13. Experimental Setup
  14. Main Results
  15. Ablation Studies
  16. ...and 3 more sections

Key Result

Theorem 4.5

Under Assumption ass1-ass3, if the semantic homogeneity $\mathcal{H}(n_i;\mathcal{A}(n_i);\tau)$ of any node $i\in\mathcal{S}$ is a $(\beta,\epsilon)$-approximation to the $\alpha(\bm{x}_{i},\hat{y}_{i},y_i)$, i.e., $|\mathcal{H}(n_i;\mathcal{A}(n_i);\tau)-\beta\cdot\alpha(\bm{x}_{i},\hat{y}_{i},y_i where $\bm{w}^{*}$ is the optimal parameters setup.

Figures (6)

  • Figure 1: Conventional methods either overlook the relational information in distinguishing noisy labels (e.g., small loss criterion), or resort to regularization for extra supervision signals that are hard to separate clean from noise nodes (e.g., consistency regularization, auxiliary tasks using edges). By comparison, this paper explicitly evaluates the reliability of nodes from carefully selected relational information in the graph topology.
  • Figure 2: The overall framework of the proposed DREAM. DREAM first selects a set of anchors relative to each target node using proximity-aware anchor selection and topology-aware anchor selection. Then, the semantic homogeneity score is computed according to the relation between the anchors and the target node. Finally, the semantic homogeneity score is used as an indicator of the reliability of each labeled node to reweight the optimization objective dynamically during optimization.
  • Figure 3: Hyperparameter analysis of DREAM on three datasets (i.e., Cora, CiteSeer, and PubMed). We show the accuracy under different numbers of proximity-aware anchors (i.e., $k_P$, left) and topology-aware anchors (i.e., $k_T$, right).
  • Figure 4: Evolution of semantic homogeneity scores of clean and noisy nodes on Cora and CiteSeer under 30% uniform label noise. DREAM successfully separates clean from noisy nodes.
  • Figure 5: Performance under different levels of uniform label noise on (a) Cora and (b) CiteSeer. The accuracy of the proposed DREAM degrades significantly slower than baseline methods as the noise rate increases from $0$ to $0.5$.
  • ...and 1 more figures

Theorems & Definitions (3)

  • Remark 4.3
  • Theorem 4.5: Proof is deferred to Appendix
  • Remark 4.6