Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A Survey of Pretraining on Graphs: Taxonomy, Methods, and Applications (2202.07893v2)

Published 16 Feb 2022 in cs.LG, cs.SI, and q-bio.BM

Abstract: Pretrained LLMs (PLMs) such as BERT have revolutionized the landscape of NLP. Inspired by their proliferation, tremendous efforts have been devoted to Pretrained Graph Models (PGMs). Owing to the powerful model architectures of PGMs, abundant knowledge from massive labeled and unlabeled graph data can be captured. The knowledge implicitly encoded in model parameters can benefit various downstream tasks and help to alleviate several fundamental issues of learning on graphs. In this paper, we provide the first comprehensive survey for PGMs. We firstly present the limitations of graph representation learning and thus introduce the motivation for graph pre-training. Then, we systematically categorize existing PGMs based on a taxonomy from four different perspectives. Next, we present the applications of PGMs in social recommendation and drug discovery. Finally, we outline several promising research directions that can serve as a guideline for future research.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Jun Xia (76 papers)
  2. Yanqiao Zhu (45 papers)
  3. Yuanqi Du (52 papers)
  4. Stan Z. Li (222 papers)
Citations (39)

Summary

A Survey of Pre-training on Graphs: Taxonomy, Methods and Applications

The paper "A Survey of Pre-training on Graphs: Taxonomy, Methods and Applications" provides a detailed overview of the evolving field of pre-trained graph models (PGMs). The authors present a structured taxonomy of PGMs, discuss various model architectures, and evaluate the implications these models have on different applications, particularly in domains like social recommendations and drug discovery.

Limitations of Graph Representation Learning

The paper begins by discussing the limitations of traditional graph representation learning, emphasizing challenges such as scarce labeled data and poor out-of-distribution generalization. These issues hinder the adaptation of Graph Neural Networks (GNNs) to novel or varied tasks. To address these challenges, the paper draws a parallel to the NLP advancements brought by pre-trained LLMs, advocating for similar pre-training paradigms in the graph domain.

Taxonomy and Model Architectures

The survey categorizes existing PGMs based on their historical development, model architectures, pre-training strategies, tuning strategies, and applications. It delineates two primary generations of PGMs:

  1. First-Generation PGMs: Focused on pre-trained graph embeddings, akin to learning representations for basic graph tasks such as node clustering and link prediction.
  2. Second-Generation PGMs: Aim to pre-train a versatile encoder capable of adaptation across diverse tasks, integrating architectures like GNNs and Transformer hybrids to optimize transfer learning.

Model architectures are divided into standalone Graph Neural Networks (like GIN) and hybrid models combining GNNs with Transformers to leverage their high expressive potential.

Pre-training Strategies

The authors detail various pre-training strategies categorized into supervised and unsupervised techniques. These include:

  • Supervised Strategies: Utilize labeled data, such as molecular properties, to pre-train GNNs.
  • Unsupervised Strategies: Exploit methods like Graph AutoEncoders (GAEs) for reconstruction tasks and Graph Contrastive Learning (GCL) techniques, such as Instance Discrimination and Deep InfoMax, to capture inherent graph structures without explicit supervision.

The survey also explores knowledge-enriched pre-training as an extension, incorporating domain-specific information to enhance model utility.

Tuning Strategies and Applications

The paper highlights several challenges with fine-tuning pre-trained models, such as overfitting and catastrophic forgetting. Strategies like meta learning and regularization techniques are suggested to address these issues.

Applications of PGMs extend into areas like social network recommendations, where they mitigate issues like cold-start, and drug discovery, where they accelerate tasks like molecular property prediction and drug interaction simulation.

Implications and Future Directions

This paper emphasizes the potential of PGMs to transform complex graph-based domains through improved representation learning and transferability. Future research directions include improving knowledge transfer methods, enhancing model interpretability, and expanding PGMs' applications into more specialized biochemical tasks. The survey invites further exploration into how insights from PGMs can be applied to macromolecule analysis, potentially unlocking broader scope and efficacy in scientific research.

In conclusion, this comprehensive survey serves as a foundational resource for researchers interested in exploring the capabilities and applications of pre-trained graph models, offering a systematic overview and identifying avenues for future inquiry.