MPrompt: A Pretraining-Prompting Scheme for Enhanced Fewshot Subgraph Classification
Name
xu-muhuaxu-meng-eecs-2024-thesis.pdf
Description
Thesis PDF
Size
1.04 MB
Format
Adobe PDF
Checksum (MD5)
c5a3370f1e2dfea063b91363c2184e57
Author(s)
Xu, Muhua
Advisor(s)
Arvind
Chen, Jie
Chen, Xuhao
Date Issued
May 2024
Publisher
Massachusetts Institute of Technology
Abstract
Motivated by the significant progress in NLP prompt learning, there have been great research interests recently in adopting the prompting mechanism for graph machine learning. Despite the prior success of prompting methods applied in node-level and graph-level learning tasks, subgraph-level tasks are highly underexplored, and the potential of prompting remains unclear. This thesis fills this gap by exploring the prompting mechanism for subgraph classification, which is a much more challenging task as it requires understanding both global and local graph structures. In this work, we build upon state-of-the-art self-supervised graph learning models to develop a subgraph-specific prompting scheme Membership Prompt (MPrompt) based on traditional graph neural networks (GNN). Our proposed prompting scheme relies on node membership knowledge to help GNN distinguish between border and local connections, which increases its expressive power while maintaining the prompt’s independence from any specific dataset or model architecture. Additionally, we also present Subgraph Reconstructive Pretraining (SRP) which can provide MPrompt with better structural embeddings during pretraining. Experiments are conducted on both synthetic and real-world datasets, including protein function prediction and social network analysis. Our method demonstrated performance improvement under few-shot experiment setting and maintained comparable performance in full-shot settings while requiring less computation.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
