Research map / Image generation / Fast diffusion sampling

Learning What Not to Impute: An Uncertainty-Aware Diffusion Framework for Meaningful Missingness

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion models as perception tools · 0 citations (OpenAlex)

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 Diff-Joint, an iterative diffusion framework that jointly models tabular values and a meaningful-missingness mask, separating intrinsically missing entries from observation-induced missing ones via uncertainty-aware aggregation of conditional samples.

Problem

Standard imputation treats all missing entries as recoverable values, but some missingness is semantically meaningful (intrinsically absent) and should be preserved; the two sources are indistinguishable from observed data alone.

Approach

Augment the diffusion state with a binary meaningful-missingness mask and alternate between diffusion model refitting and aggregation of K conditional samples, labeling missing entries as meaningful when they fall in a high-uncertainty k-means cluster or have majority-vote mask samples equal to 1.

Key results

• On the Bayesian network dataset under MCAR at 10% missing ratio, Diff-Joint reaches 78.34% discrete ACC versus 35.21% for the best baseline CMAE.
• MM recall on the Bayesian network dataset ranges from 92.02% (10% ratio) to 76.99% (40%), with precision falling from 73.45% to 43.54%.
• On MIMIC-IV-ED at 10% MCAR, Diff-Joint achieves 72.00% discrete accuracy versus 63.85% for CACTI, while continuous RMSE is close to the best (8.31 vs 8.24).

Surprising findings

• On continuous columns, Diff-Joint is not always the best point imputer; baselines such as CMAE achieve lower RMSE, attributed to the joint objective balancing MM identification.

Limitations

• Assumes known MM candidate columns and a detectable uncertainty gap between MM and observation-induced entries.
• Iterative diffusion procedure is more computationally expensive than non-iterative baselines.
• Identifiability results rely on restrictive regimes (known MCAR probabilities or known subset of MM columns).

Experimental setup

Datasets: synthetic Bayesian-network dataset MIMIC-IV-ED

Benchmarks: synthetic Bayesian-network dataset MIMIC-IV-ED

Metrics: precision recall ACC RMSE MAE Macro-F1 ROC-AUC Weighted-F1

Baselines: Mean/Mode missForest CACTI (CMAE) DiffPuter

Builds on: DiffPuter score-based diffusion models missForest

Closest papers by meaning

• Rethinking the Diffusion Models for Missing Data Imputation: A Gradient Flow Perspective (NeurIPS 2024 · similarity 0.90)
• DiffPuter: Empowering Diffusion Models for Missing Data Imputation (ICLR 2025 · similarity 0.89)
• Distribution Shift in Missing Data Imputation: A Risk-Based Perspective and Importance-Weighted Correction under MAR (NeurIPS 2026 · similarity 0.89)
• Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification (NeurIPS 2025 · similarity 0.88)
• CACTI: Leveraging Copy Masking and Contextual Information to Improve Tabular Data Imputation (ICML 2025 · similarity 0.88)
• Missing Data Imputation by Reducing Mutual Information with Rectified Flows (NeurIPS 2025 · similarity 0.87)