Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

E2EG: End-to-End Node Classification Using Graph Topology and Text-based Node Attributes

Tu Anh Dinh, Jeroen den Boef, Joran Cornelisse, Paul Groth

TL;DR

Textual node classification benefits from both node text and graph topology, but two-stage pipelines risk information loss and added complexity. E2EG unifies text encoding with graph structure in an end-to-end, multi-task framework that jointly optimizes node classification and neighborhood prediction, enabling lighter encoders and tighter integration. The approach yields competitive accuracy with fewer parameters on ogbn-arxiv and ogbn-products in both transductive and inductive settings, and qualitative analysis shows improved use of topological context. Overall, E2EG offers a practical, compact alternative to GIANT-based pipelines with real-world deployment advantages.

Abstract

Node classification utilizing text-based node attributes has many real-world applications, ranging from prediction of paper topics in academic citation graphs to classification of user characteristics in social media networks. State-of-the-art node classification frameworks, such as GIANT, use a two-stage pipeline: first embedding the text attributes of graph nodes then feeding the resulting embeddings into a node classification model. In this paper, we eliminate these two stages and develop an end-to-end node classification model that builds upon GIANT, called End-to-End-GIANT (E2EG). The tandem utilization of a main and an auxiliary classification objectives in our approach results in a more robust model, enabling the BERT backbone to be switched out for a distilled encoder with a 25% - 40% reduction in the number of parameters. Moreover, the model's end-to-end nature increases ease of use, as it avoids the need of chaining multiple models for node classification. Compared to a GIANT+MLP baseline on the ogbn-arxiv and ogbn-products datasets, E2EG obtains slightly better accuracy in the transductive setting (+0.5%), while reducing model training time by up to 40%. Our model is also applicable in the inductive setting, outperforming GIANT+MLP by up to +2.23%.

E2EG: End-to-End Node Classification Using Graph Topology and Text-based Node Attributes

TL;DR

Textual node classification benefits from both node text and graph topology, but two-stage pipelines risk information loss and added complexity. E2EG unifies text encoding with graph structure in an end-to-end, multi-task framework that jointly optimizes node classification and neighborhood prediction, enabling lighter encoders and tighter integration. The approach yields competitive accuracy with fewer parameters on ogbn-arxiv and ogbn-products in both transductive and inductive settings, and qualitative analysis shows improved use of topological context. Overall, E2EG offers a practical, compact alternative to GIANT-based pipelines with real-world deployment advantages.

Abstract

Node classification utilizing text-based node attributes has many real-world applications, ranging from prediction of paper topics in academic citation graphs to classification of user characteristics in social media networks. State-of-the-art node classification frameworks, such as GIANT, use a two-stage pipeline: first embedding the text attributes of graph nodes then feeding the resulting embeddings into a node classification model. In this paper, we eliminate these two stages and develop an end-to-end node classification model that builds upon GIANT, called End-to-End-GIANT (E2EG). The tandem utilization of a main and an auxiliary classification objectives in our approach results in a more robust model, enabling the BERT backbone to be switched out for a distilled encoder with a 25% - 40% reduction in the number of parameters. Moreover, the model's end-to-end nature increases ease of use, as it avoids the need of chaining multiple models for node classification. Compared to a GIANT+MLP baseline on the ogbn-arxiv and ogbn-products datasets, E2EG obtains slightly better accuracy in the transductive setting (+0.5%), while reducing model training time by up to 40%. Our model is also applicable in the inductive setting, outperforming GIANT+MLP by up to +2.23%.
Paper Structure (20 sections, 2 figures, 6 tables)

This paper contains 20 sections, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Background and related work
  3. Graph
  4. Machine learning on graphs
  5. Textual node classification
  6. The GIANT framework
  7. Methodology
  8. Main task delay
  9. Additional fine-tuning round
  10. Transductive data utilization
  11. Encoders
  12. Experimental Setup
  13. Results
  14. Results
  15. Effect of proposed training strategies
  16. ...and 5 more sections

Figures (2)

  • Figure 1:
  • Figure 2: Samples from ogbn-arxiv. The predicted node is the big circle at the top right. The other circles represent the two-hop neighborhood of the predicted node. The big node’s color represents the predicted class, the other nodes' colors represent the actual classes. The text at the bottom is the predicted node's raw text.