A Subgraph-Based Framework for Self-Supervised Graph Representation Learning

Self-supervised learning (SSL) has transformed fields such as computer vision and natural language processing, but extending its success to graph-structured data poses unique challenges. Unlike images or text, graphs lack natural tokenization, exhibit sparsity, and require permutation equivariance, making direct adaptation of existing SSL methods non-trivial.

In this work, we present a novel framework for self-supervised pre-training on graphs that avoids costly graph augmentations and instead leverages the rich and expressive nature of the latent representation space. Our method is inspired by the Joint Embedding Predictive Architecture (JEPA), which we adapt to the graph domain. Specifically, we partition each graph into disjoint subgraphs, processing one subset with a context encoder and the other with a target encoder. We aim to contribute a simple pretext task and a generalizable architecture applicable to graphs across diverse domains.