An Analysis of Role Semantics in Directed Dynamic Graph Modeling for Edge Classification

Abstract

Edge classification on directed dynamic graphs is critical for understanding evolving user-item interactions and behavioral patterns in recommender systems and social platforms. Existing dynamic graph architectures typically process these interactions using shared parameters for nodes with semantically distinct roles, including source, destination, query, and neighbor. While abundant supervision in tasks such as link prediction provides sufficient gradient signal for models to implicitly deduce role semantics and overcome this conflation, there are often far fewer labels available for edge classification. We hypothesize that in such regimes, the ability of models to implicitly separate roles is restricted, degrading downstream performance. To empirically investigate this relationship, we introduce an experimental framework for sequence-based models that explicitly disentangles functional roles using targeted embeddings and an asymmetric pretraining objective. By providing architectural priors for these roles, the framework serves as an evaluative vehicle to isolate and measure the resulting performance delta across a range of label budgets. Our empirical evaluation demonstrates that this explicit role disentanglement yields substantial and consistent relative performance gains over baselines. Furthermore, representational analyses reveal that the mechanism by which the framework provides these benefits is a bifurcation of source and destination latent spaces that improves linear separability.