Research map / Probabilistic and scientific modeling / Dynamical systems
Guided Structural Inference: Leveraging Priors with Soft Gating Mechanisms
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 SGSI, a VAE-based structural inference method that injects partial prior knowledge (known edges, sparsity, node degrees) via soft gating, clone-clamping of known edges, and excluding them from the KL term.
Problem
Latent relational structure inference from time series cannot easily integrate partial prior knowledge (known present/absent edges, node-degree constraints, global sparsity) without destabilizing training or conflicting with VAE/IB assumptions.
Approach
Use learnable sigmoid edge gates that multiply messages, clone the gates and clamp known edges to 1 or 0 in the forward pass only, drop known edges from the KL term, and add soft sparsity and degree penalties instead of hard constraints.
Key results
- SGSI with sparsity constraint reaches 96.90% AUROC on VN_SP_30 versus 92.76% for SGSI without constraints, with 2-6% gains from sparsity.
- SGSI achieves up to 9% delta AUROC improvement with partial prior knowledge, while baselines NRI, MPM, ACD, iSIDG, RCSI show negative delta AUROC at high known-edge fractions.
- Ablations show removing KL skipping for known edges or removing cloning yields higher loss and instability relative to default SGSI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global sparsity constraint (ablation) | SGSI · VN SP 30 | AUROC: 92.76% to 96.90% | Helped |
| add in-degree constraints (ablation) | SGSI · VN SP 50 | AUROC: 90.92% to 95.90% | Helped |
| add both in-degree and out-degree constraints (ablation) | SGSI · VN NS 50 | AUROC: 79.05% to 83.56% | Helped |
| skip KL regularization for known, clamped edges (ablation) | SGSI · VN SP 50 | adjacency accuracy | Helped |
| clone gating variables before clamping (ablation) | SGSI · VN SP 50 | adjacency accuracy | Helped |
| use SGSI with sparsity rather than unconstrained SGSI (ablation) | SGSI · VN SP 50 | AUROC: 90.92 -> 96.43 | Helped |
| use SGSI with both degree constraints rather than unconstrained SGSI (ablation) | SGSI · VN SP 100 | AUROC: 88.17 -> 94.81 | Helped |
| use SGSI with global sparsity rather than unconstrained SGSI (ablation) | SGSI · VN NS 100 | AUROC: 77.90 -> 81.54 | Helped |
| use SGSI with sparsity rather than unconstrained SGSI (ablation) | SGSI · VN SP 15 | AUROC: 95.29 -> 97.80 | Helped |
| use SGSI with both degree constraints rather than unconstrained SGSI (ablation) | SGSI · VN NS 30 | AUROC: 88.06 -> 95.00 | Helped |
Surprising findings
- Naive overwriting of latent edges with known values makes most VAE baselines (NRI, MPM, ACD, iSIDG, RCSI) show negative AUROC gains when more edges are known.
Limitations
- Baselines cannot natively integrate sparsity or degree constraints, so comparisons under those constraints are limited to SGSI variants.
- Hyperparameters (beta, lambda_sparsity, lambda_deg) require grid or Bayesian search per dataset.
- Robustness experiments with noisy priors are described as preliminary.
Experimental setup
Datasets: Springs Simulations NetSims Synthetic biological networks (LI, LL, CY, BF, TF, BF-CV) via BoolODE StructInfer Benchmark Vascular Networks (VN) PEMS03 PEMS04 PEMS07
Benchmarks: Springs Simulations NetSims BF-CV VN_SP_15 VN_SP_30 VN_SP_50 VN_SP_100 VN_NS_15 VN_NS_30 VN_NS_50 VN_NS_100
Metrics: AUROC delta AUROC
Baselines: NRI MPM ACD iSIDG RCSI ALaSI SICSM
Builds on: NRI information bottleneck ALaSI SICSM
Closest papers by meaning
- Benchmarking Structural Inference Methods for Interacting Dynamical Systems with Synthetic Data (NeurIPS 2024 · similarity 0.87)
- ProDAG: Projected Variational Inference for Directed Acyclic Graphs (NeurIPS 2025 · similarity 0.86)
- A Graph Dynamics Prior for Relational Inference (AAAI 2024 · similarity 0.86)
- Interpretable Relational Inference with LLM-Guided Symbolic Dynamics Modeling (NeurIPS 2026 · similarity 0.85)
- Graph Structure Inference with BAM: Neural Dependency Processing via Bilinear Attention (NeurIPS 2024 · similarity 0.85)
- TreeVI: Reparameterizable Tree-structured Variational Inference for Instance-level Correlation Capturing (NeurIPS 2024 · similarity 0.85)
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.