Graph Neural ODEs combine Graph Neural Networks (GNNs) with Neural ODEs — defining continuous-time dynamics on graph-structured data where node features evolve according to an ODE parameterized by a GNN, enabling continuous-depth message passing and diffusion on graphs.

How Graph Neural ODEs Work

• Graph Input: A graph with node features $h_i(0)$ at time $t=0$.
• Continuous Dynamics: $frac{dh_i}{dt} = f_ heta(h_i, {h_j : j in N(i)}, t)$ — node features evolve based on local neighborhood.
• ODE Solver: Integrate the dynamics from $t=0$ to $T$ using an adaptive ODE solver.
• Output: Node features at time $T$ are used for classification, regression, or generation.

Why It Matters

• Over-Smoothing: Continuous dynamics with adaptive depth naturally addresses the over-smoothing problem of deep GNNs.
• Continuous Depth: No fixed number of message-passing layers — depth adapts to the task and graph structure.
• Physical Systems: Natural model for physical processes on networks (heat diffusion, epidemic spreading, traffic flow).

Graph Neural ODEs are continuous GNNs — replacing discrete message-passing layers with continuous dynamics for adaptive-depth graph processing.