Research map / Probabilistic and scientific modeling / Molecules and proteins
Amortized Network Intervention to Steer the Excitatory Point Processes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Amortized Network Interventions (ANI), pooling policies across subgraphs with a permutation-equivalent bi-contrastive embedding, using H-step lookahead model-based RL on a neural-ODE excitatory point process model with a mean-field reward approximation.
Problem
Steering excitatory point processes (event flows such as epidemic spread or traffic congestion) over dynamic graphs by intervening on edges, where the combinatorial action space and heterogeneous regional dynamics make per-region policy learning hard to scale and transfer.
Approach
Learn a shared amortized policy across subgraphs with a permutation-equivalent representation (bi-contrastive metric embeddings with a permutation equivalent metric), plan via H-step lookahead on a neural-ODE jump model, and replace rollouts with an analytical mean-field estimate of expected event counts.
Key results
- On COVID data, the adaptive plus amortized policy reaches reduced intensities of 0.47 (Georgia-0), 0.71 (Alabama-0), 0.39 (Georgia-1), and 0.54 (West Virginia-0) versus at most 0.18 for non-amortized adaptive baselines.
- Ablation on West Virginia: the PEM plus Bi-CME configuration gives 0.54 reduced intensity, while PSM variants reach only 0.02-0.05, a roughly 90% degradation.
- On synthetic MHP data, the model-based approach reportedly achieves intensity reduction comparable to SAC and PPO without direct environment interaction, while NHPI struggles on high-frequency events.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an adaptive amortized policy instead of a non-adaptive, non-amortized policy | ANI · Georgia-0 | reduced intensity: -0.05(0.18) -> 0.47(0.14) | Helped |
| Use an adaptive amortized policy instead of a non-adaptive, non-amortized policy | ANI · Alabama-0 | reduced intensity: 0.08(0.06) -> 0.71(0.42) | Helped |
| Use an adaptive amortized policy instead of a non-adaptive, non-amortized policy | ANI · Georgia-1 | reduced intensity: -0.07(0.11) -> 0.39(0.27) | Helped |
| Use an adaptive amortized policy instead of a non-adaptive, non-amortized policy | ANI · West Virginia-0 | reduced intensity: -0.02(0.05) -> 0.54(0.27) | Helped |
| Add adaptation to the amortized policy | ANI · Georgia-0, Alabama-0, Georgia-1, West Virginia-0 | reduced intensity: 0.21(0.43) -> 0.47(0.14); 0.18(0.58) -> 0.71(0.42); 0.06(0.24) -> 0.39(0.27); 0.02(0.02) -> 0.54(0.27) | Helped |
| Add amortization to the adaptive policy | ANI · Georgia-0, Alabama-0, Georgia-1, West Virginia-0 | reduced intensity: 0.18(0.19) -> 0.47(0.14); 0.14(0.22) -> 0.71(0.42); 0.02(0.13) -> 0.39(0.27); 0.15(0.10) -> 0.54(0.27) | Helped |
| Use Bi-CMEs instead of separately learned permutation and magnitude CMEs (ablation) | PEM-based policy embeddings · West Virginia-0 | reduced intensity: 0.01(0.06) -> 0.54(0.27); 0.05(0.02) -> 0.54(0.27) | Helped |
| Use PEM instead of PSM with Bi-CMEs (ablation) | Bi-CMEs policy embeddings · West Virginia-0 | reduced intensity: 0.05(0.02) -> 0.54(0.27) | Helped |
| Use Bi-CMEs instead of separately learned permutation and magnitude CMEs (ablation) | PSM-based policy embeddings · West Virginia-0 | reduced intensity: 0.02(0.04); 0.04(0.02); 0.05(0.02) | No clear effect |
| Use Neural Jump ODE instead of Neural Process | Synthetic 5-dimensional MHP · Synthetic 5-dimensional MHP | log-likelihood: -180 -> -33 | Helped |
Surprising findings
- Policy similarity metric (PSM)-based embeddings degrade performance by about 90% versus the proposed permutation-equivalent metric.
- The non-adaptive, non-amortized policy yields near-zero or negative reduced intensities on several COVID splits.
Limitations
- Assumes a receding horizon window approximates the optimal policy, which may degrade under long-range correlations.
- Theoretical error bound relies on strong assumptions such as Lipschitz constants below 1 and spectral radius of the influence matrix below 1.
Experimental setup
Datasets: synthetic multivariate Hawkes processes SUMO traffic simulation data COVID county case data
Benchmarks: synthetic multivariate Hawkes processes COVID county data (Georgia, Alabama, West Virginia splits) SUMO traffic intersections
Metrics: reduced intensity intensity cost log-likelihood
Baselines: SAC PPO Neural Hawkes Process Intervention (NHPI) PSM
Builds on: Neural Jump SDEs Policy Similarity Metric (PSM) SimCLR SAC PPO
Closest papers by meaning
- Dynamic Treatment on Networks (NeurIPS 2026 · similarity 0.86)
- Amortized Active Causal Induction with Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- MOCHA: Discovering Multi-Order Dynamic Causal Structure in Temporal Point Processes (NeurIPS 2026 · similarity 0.85)
- Action abstractions for amortized sampling (ICLR 2025 · similarity 0.84)
- Towards Effective Planning Strategies for Dynamic Opinion Networks (NeurIPS 2024 · similarity 0.84)
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.