Adaptive Memory Retention in Dynamic Graphs

We introduce LAMP (Long-range Adaptive Memory Propagation), a differential-equation GNN for snapshot-based dynamic graphs that combines impulsive neural ODEs with an antisymmetric parameterization and a learned, adaptive dissipation term, so the model learns how much information to conserve and how much to dissipate over space and time.

• Interactive demo
• Code on GitHub
• Paper ICML page
• OpenReview

Download Paper