TI-ODE: A New Approach for Dynamic Graph Representation Learning
Researchers propose Time-varying Interaction Graph Ordinary Differential Equations (TI-ODE) to address limitations in existing graph neural ODEs for dynamic graph representation learning. Current graph neural ODEs use a unified message passing mechanism, assuming inter-node interactions share the same function at any time, which fails to capture the diversity and time-varying nature of interaction patterns. TI-ODE decomposes the evolution function into a set of learnable interaction basis functions, each corresponding to a distinct type of inter-node interaction, combined through time-dependent learnable weights. The method is detailed in arXiv preprint 2604.24811.
Key facts
- arXiv:2604.24811 introduces TI-ODE
- TI-ODE addresses limitations of existing graph neural ODEs
- Existing methods use a unified message passing mechanism
- TI-ODE decomposes evolution function into learnable interaction basis functions
- Each basis function corresponds to a distinct interaction type
- Basis functions are combined via time-dependent learnable weights
- The work focuses on dynamic graph representation learning
- The paper is a cross-type announcement on arXiv
Entities
Institutions
- arXiv