Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Robust Simulation-Based Inference under Missing Data via Neural 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 RISE, an amortized SBI method that jointly learns a latent neural process imputation model and an NPE inference network to reduce posterior bias under missing data (MCAR/MAR/MNAR).
Problem
Simulation-based inference requires fully observed data, yet real datasets have missing values; naive imputation (zero/mean) biases the SBI posterior.
Approach
Jointly train a latent neural process imputation model with a neural posterior estimator via a variational lower bound, so the imputation distribution and inference network are learned together and the missingness mechanism (MCAR/MAR/MNAR) is modeled; a meta-learning variant (RISE-Meta) amortizes over unknown missingness proportions.
Key results
- Under MCAR at 60% missingness on GLU, RISE reaches C2ST 0.93 vs 0.96-0.97 for baselines, and RISE has the best NLPP on most Ricker/OUP settings.
- On bioactivity imputation, RISE achieves R^2 0.67 on Adrenergic and 0.26 on Kinase, above CNP (0.65, 0.24).
- Joint training beats separate training: RISE RMSE 0.41 vs 0.43 (RISE-Sep) on GLU at 10% missingness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RISE rather than the strongest reported baseline | SBI posterior estimation · GLU and GLM, MCAR and MNAR, 60% missing | C2ST: 0.96 -> 0.92 (GLU MNAR); 0.98 -> 0.97 (GLM MCAR) | Helped |
| Use RISE rather than the strongest reported baseline | SBI posterior estimation · Ricker and OUP, MCAR, 60% missing | C2ST: 0.97 -> 0.94 (Ricker); 0.94 -> 0.93 (OUP) | Helped |
| Use RISE rather than the strongest reported baseline | SBI posterior estimation · Ricker and OUP, MNAR, 60% missing | C2ST: 0.99 -> 0.95 (Ricker); 0.99 -> 0.97 (OUP) | Mixed |
| Use the neural-process imputation model in RISE rather than the strongest listed imputation baseline | RISE imputation · Adrenergic | R²: 0.65 ± 0.04 -> 0.67 ± 0.03 | Helped |
| Use the neural-process imputation model in RISE rather than the strongest listed imputation baseline | RISE imputation · Kinase | R²: 0.24 ± 0.02 -> 0.26 ± 0.03 | Helped |
| Jointly train imputation and inference rather than train them separately (ablation) | RISE · GLM, GLU at 10%, 25%, and 60% missing | RMSE: 0.67 -> 0.65; 0.99 -> 0.93; 1.31 -> 1.27 (GLM); 0.43 -> 0.41; 0.45 -> 0.43; 0.58 -> 0.56 (GLU) | Helped |
Experimental setup
Datasets: Ricker Ornstein-Uhlenbeck process GLM GLU Hodgkin-Huxley Adrenergic assays Kinase assays
Benchmarks: Ricker OUP GLM GLU Hodgkin-Huxley Adrenergic Kinase
Metrics: NLPP C2ST MMD RMSE R^2 expected coverage
Baselines: NPE-Zero NPE-Mean NPE-NN Wang et al. Simformer RISE-Sep NPE-RF-Sep QSAR CMF DNN NP Conduit
Builds on: neural posterior estimation latent neural processes Wang et al. missing-data SBI with masks Simformer
Closest papers by meaning
- Effortless, Simulation-Efficient Bayesian Inference using Tabular Foundation Models (NeurIPS 2025 · similarity 0.88)
- Simultaneous identification of models and parameters of scientific simulators (ICML 2024 · similarity 0.87)
- Consistency Models for Scalable and Fast Simulation-Based Inference (NeurIPS 2024 · similarity 0.86)
- Multilevel neural simulation-based inference (NeurIPS 2025 · similarity 0.86)
- Compositional simulation-based inference for time series (ICLR 2025 · similarity 0.86)
- Addressing Misspecification in Simulation-based Inference through Data-driven Calibration (ICML 2025 · similarity 0.86)
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.