MoJEPA: Joint-Embedding Predictive Learning for Dynamic Graphs with Discrete Morse Theory

Self-supervised pretraining on dynamic graphs can reduce reliance on task-specific labels and support transfer across datasets. In joint-embedding predictive architectures (JEPAs), the learning task is jointly defined by what is masked and what context remains available for prediction. On dynamic graphs, these choices face two fundamental coupled challenges. First, masking interactions also removes relational evidence needed for prediction. Second, changing interactions limit the context available at each time, leaving useful relations observed at other times unavailable for prediction. To address these challenges, we design _MoJEPA_ which combines partial interaction masking with a shared Morse reference forest. We partially mask interactions so that the model can predict missing evidence from the relations that remain visible. To enrich this predictive context, we introduce _Morse reference context_, which leverages discrete Morse theory from low-dimensional topology to reduce the relations observed across a window to a sparse rooted forest. This allows the relations observed at other times to support information exchange at the target time. A relation observed at one time can subsequently support the exchange of node states at another, even when the corresponding interaction is absent. This combination allows prediction to use context beyond the interactions present at an individual time. Our experiments demonstrate that MoJEPA leads on all 9 within-dataset tasks and 15 of 18 transfer directions, with relative gains over the strongest external baseline up to 13.13% and 12.27% for within-dataset prediction and transfer respectively.