Homological Representation Learning for Molecular Graphs

Yoshihiro Maruyama, Arisa Yasuda

Keywords: representation learning; homology; algebraic topology; molecular graph

TL;DR: Algebraic topology provides a regularization method for representation learning

Abstract: We propose Homological Representation Learning (HomRL), an architecture-agnostic regularization method for graph encoders that aligns latent embeddings with an efficiently computable homological signature of the input. In this paper, we give both theoretical results on representation invariance bounds and empirical results on molecular graph classification tasks.

Submission Number: 144