Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weighted Graph Structure Learning with Attention Denoising for Node Classification

Tingting Wang, Jiaxin Su, Haobing Liu, Ruobing Jiang

TL;DR

This work addresses robust node classification on noisy weighted graphs by introducing Edge Weight-aware Graph Structure Learning (EWGSL), which couples edge-weighted attention with sparse graph structure learning and a denoising-aware loss. The method redefines GAT attention to incorporate edge weights and uses alpha-entmax to prune irrelevant edges, while a joint L = L_C + \eta L_I term (InfoNCE) leverages denoised graph weights for contrastive regularization. Key contributions include jointly learning edge weights and graph structure, integrating an adaptive sparsification mechanism, and achieving substantial improvements in Micro-F1 (average 17.8% over baselines) and ACC in diverse noisy datasets. The proposed approach enhances robustness to noise in weighted graphs and holds promise for scalable, real-world graph learning tasks.

Abstract

Node classification in graphs aims to predict the categories of unlabeled nodes by utilizing a small set of labeled nodes. However, weighted graphs often contain noisy edges and anomalous edge weights, which can distort fine-grained relationships between nodes and hinder accurate classification. We propose the Edge Weight-aware Graph Structure Learning (EWGSL) method, which combines weight learning and graph structure learning to address these issues. EWGSL improves node classification by redefining attention coefficients in graph attention networks to incorporate node features and edge weights. It also applies graph structure learning to sparsify attention coefficients and uses a modified InfoNCE loss function to enhance performance by adapting to denoised graph weights. Extensive experimental results show that EWGSL has an average Micro-F1 improvement of 17.8% compared with the best baseline.

Weighted Graph Structure Learning with Attention Denoising for Node Classification

TL;DR

This work addresses robust node classification on noisy weighted graphs by introducing Edge Weight-aware Graph Structure Learning (EWGSL), which couples edge-weighted attention with sparse graph structure learning and a denoising-aware loss. The method redefines GAT attention to incorporate edge weights and uses alpha-entmax to prune irrelevant edges, while a joint L = L_C + \eta L_I term (InfoNCE) leverages denoised graph weights for contrastive regularization. Key contributions include jointly learning edge weights and graph structure, integrating an adaptive sparsification mechanism, and achieving substantial improvements in Micro-F1 (average 17.8% over baselines) and ACC in diverse noisy datasets. The proposed approach enhances robustness to noise in weighted graphs and holds promise for scalable, real-world graph learning tasks.

Abstract

Node classification in graphs aims to predict the categories of unlabeled nodes by utilizing a small set of labeled nodes. However, weighted graphs often contain noisy edges and anomalous edge weights, which can distort fine-grained relationships between nodes and hinder accurate classification. We propose the Edge Weight-aware Graph Structure Learning (EWGSL) method, which combines weight learning and graph structure learning to address these issues. EWGSL improves node classification by redefining attention coefficients in graph attention networks to incorporate node features and edge weights. It also applies graph structure learning to sparsify attention coefficients and uses a modified InfoNCE loss function to enhance performance by adapting to denoised graph weights. Extensive experimental results show that EWGSL has an average Micro-F1 improvement of 17.8% compared with the best baseline.

Paper Structure

This paper contains 15 sections, 15 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Weight Learning Methods
  4. Structure Learning Methods
  5. The Proposed Model
  6. Problem Definition
  7. Edge-aware Weight Learning
  8. Graph Sparsity Structure Learning
  9. Loss Function
  10. Experiments
  11. Experiment Settings
  12. Experment Results
  13. Ablation Study
  14. Case Study
  15. Conclusion

Figures (3)

  • Figure 1: The architecture of EWGSL. The input consists of a weighted graph $G$ and a set of partially labeled nodes $Y_L$. Within the multiple layers of the EWGSL structure, as shown in the green box, node representations are progressively learned. Specifically, EWGSL incorporates edge weight information when calculating attention scores and employs the $\alpha$-entmax function for normalization. This process accurately captures the relationships between nodes by removing edges with low attention and reallocating attention to relevant nodes. This process is repeated multiple times (three times), ultimately using multi-head attention to learn node representations. The final loss function comprises two parts: $L_C$ and $L_I$.
  • Figure 2: Abiation study results.
  • Figure 3: Case study results.