SUREL+: From Walks to Sets for Scalable Subgraph-based Graph Representation Learning

Explore a 50-minute conference talk on SUREL+, an innovative framework for scalable subgraph-based graph representation learning. Dive into the background of subgraph-based graph representation learning (SGRL) and its challenges. Learn about SUREL, a method that accelerates SGRL by sampling random walks offline. Discover how SUREL+ improves upon SUREL by using node sets instead of walks, resulting in significant speedups and improved prediction accuracy. Examine experimental settings, results, and engage in Q&A interludes throughout the presentation. Investigate how SUREL+ can boost Message Passing Neural Networks (MPNNs) with Bloom Signatures. Conclude with a summary of findings and their implications for AI in drug discovery and other graph-based applications.

- Intro
- Background
- SUREL
- Experimental Settings + Results
- Q&A Interlude
- SUREL+
- Experimental Settings + Results
- Q&A Interlude
- Boost MPNNs with Bloom Signatures
- Results
- Summary + Conclusions