A research team says treating the two sides of a directed network edge as genuinely different things — not just two instances of the same node type — measurably improves edge classification on dynamic graphs.
The paper introduces DyGnROLE, a Transformer-based model that gives source and destination nodes separate embedding tables and role-specific positional encodings. Most existing dynamic graph architectures share parameters across both ends of an edge, which the authors argue obscures the asymmetric behavioral patterns that distinguish, say, a fraudster from a victim, or a message sender from a recipient. To address the chronic shortage of labeled edges in real-world graph datasets, the team also devised a self-supervised pretraining method called Directional Role Alignment, which trains source representations to retrieve their corresponding destination representations while blocking previously seen pairs from being reused as negatives — injecting a temporal direction into the learning signal. Testing across four edge classification tasks and eight datasets, DyGnROLE consistently beat a range of state-of-the-art baselines.
Edge classification on dynamic graphs matters in fraud detection, network security, and recommendation systems — domains where who initiated an interaction is as important as the interaction itself. The finding that shared-parameter architectures have been systematically underperforming on directed graphs suggests a quiet ceiling on a class of models deployed in production today.
The paper does not name the datasets beyond counting eight of them, so independent replication will be the real test — benchmark cherry-picking is a longstanding concern in graph machine learning.