Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Global-Local Graph Neural Networks for Node-Classification

Moshe Eliasof, Eran Treister

TL;DR

GLGNN addresses node classification by jointly leveraging local node features and learnable global label features. It learns a distinct feature vector per class via a global readout and per-label expansions, and fuses these with the node representations to form a node-classification map through a per-label dot product followed by SoftMax. The method minimizes a joint loss combining cross-entropy with a global-local loss that pulls label features toward corresponding nodes and pushes apart non-corresponding ones, with a clamp parameter to control out-of-class distances. Across semi- and fully-supervised settings and multiple backbones (GCN, GAT, GCNII), GLGNN yields consistent improvements on standard benchmarks and graph classification tasks, while maintaining modest computational overhead. This demonstrates the value of explicit label-driven global information for graph-based classification and suggests avenues for further integration of global cues and pseudo-label techniques.

Abstract

The task of graph node classification is often approached by utilizing a local Graph Neural Network (GNN), that learns only local information from the node input features and their adjacency. In this paper, we propose to improve the performance of node classification GNNs by utilizing both global and local information, specifically by learning label- and node- features. We therefore call our method Global-Local-GNN (GLGNN). To learn proper label features, for each label, we maximize the similarity between its features and nodes features that belong to the label, while maximizing the distance between nodes that do not belong to the considered label. We then use the learnt label features to predict the node classification map. We demonstrate our GLGNN using three different GNN backbones, and show that our approach improves baseline performance, revealing the importance of global information utilization for node classification.

Global-Local Graph Neural Networks for Node-Classification

TL;DR

GLGNN addresses node classification by jointly leveraging local node features and learnable global label features. It learns a distinct feature vector per class via a global readout and per-label expansions, and fuses these with the node representations to form a node-classification map through a per-label dot product followed by SoftMax. The method minimizes a joint loss combining cross-entropy with a global-local loss that pulls label features toward corresponding nodes and pushes apart non-corresponding ones, with a clamp parameter to control out-of-class distances. Across semi- and fully-supervised settings and multiple backbones (GCN, GAT, GCNII), GLGNN yields consistent improvements on standard benchmarks and graph classification tasks, while maintaining modest computational overhead. This demonstrates the value of explicit label-driven global information for graph-based classification and suggests avenues for further integration of global cues and pseudo-label techniques.

Abstract

The task of graph node classification is often approached by utilizing a local Graph Neural Network (GNN), that learns only local information from the node input features and their adjacency. In this paper, we propose to improve the performance of node classification GNNs by utilizing both global and local information, specifically by learning label- and node- features. We therefore call our method Global-Local-GNN (GLGNN). To learn proper label features, for each label, we maximize the similarity between its features and nodes features that belong to the label, while maximizing the distance between nodes that do not belong to the considered label. We then use the learnt label features to predict the node classification map. We demonstrate our GLGNN using three different GNN backbones, and show that our approach improves baseline performance, revealing the importance of global information utilization for node classification.
Paper Structure (32 sections, 13 equations, 3 figures, 10 tables)

This paper contains 32 sections, 13 equations, 3 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Graph Neural Networks
  4. Improved training of GNNs
  5. Global information in GNNs
  6. Label information in GNNs
  7. Preliminaries and notations
  8. Definitions
  9. GNN Backbones
  10. GCN
  11. GAT
  12. GCNII
  13. Method
  14. Local node features
  15. Global label features
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Our GLGNN architecture. We learn node and label features, and fuse them to obtain a response per node for each label. $\oplus$ denotes feature-wise concatenation. $\odot$ denotes matrix-vector dot product. $\rm{MLP}_q$ implements Equation \ref{['eq:SE']} for $0 \leq q \leq k-1$.
  • Figure 2: tSNE embedding of learnt label- and node-features of Cora. The similarity of the label features and the corresponding node features shows the clustering effect of incorporating global information.
  • Figure 3: Ablation study accuracy (%) on semi-supervised node classification using our GLGCN vs. (a) the channel expansion rate $\rm{e}$ from Equation \ref{['eq:SE']}. (b) the cut-off value $\rm{r}$ from Equation \ref{['eq:globalLocalLoss']}. (c) The dropout value $p$. (d) The value of $\gamma$ from Equation \ref{['eq:totalObjective']}.