Research map / Time series / Causal and generative time series

Discretizing Continuous Time Series for Imputation with Masked Diffusion Training

NeurIPS 2026 · Time series › Causal and generative time series · approach: Irregular series and imputation · 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 MDTIM, a masked diffusion model for time series imputation that discretizes values with stochastic discretization and ordinal-aware soft labels, predicting clean values with expectation-based decoding.

Problem

Time series imputation requires representations and objectives aligned with partial observability, but existing methods mix missing placeholders with observed values and continuous diffusion models predict noise rather than the original signal.

Approach

Adapts masked diffusion language model training to imputation: a [MASK] token separates missing from observed values, the model predicts clean token distributions, and stochastic discretization with truncated Gaussian soft labels preserves ordinal structure, with continuous values recovered as expectations over bins.

Key results

• On Energy at 30% uniform missing, MDTIM reaches MAE 0.044 vs 0.094 for CSDI and 0.050 for FGTI, the best of the baselines listed.
• At L=192 with 30% uniform missing, MDTIM achieves MAE 0.123 vs 0.145 for SAITS and 0.192 for BRITS.
• On Energy uniform 70%, the Small MDTIM (0.65M params, 4.05 s) gets MAE 0.092 vs CSDI Large (4.49M params, 894.76 s) at 0.128.

Surprising findings

• FGTI, a frequency-aware diffusion model, outperforms MDTIM on the synthetic Sine dataset.
• The Ordinal-Aware Soft Labeling gives no further improvement on Energy beyond expectation-based unmasking.

Limitations

• Optimal vocabulary size K varies across datasets and must be tuned manually (K=40 chosen).
• Spectral consistency loss is disabled for PhysioNet because irregular data lacks stable frequency structure.
• Evaluation limited to standard benchmarks of moderate length; long sequences and non-stationary domains left for future work.

Experimental setup

Datasets: Energy ETTh Weather Sine PhysioNet 2012

Benchmarks: Energy ETTh Weather Sine PhysioNet 2012

Metrics: MAE MSE CRPS inference time parameter count

Baselines: BRITS MRNN GRU-D Transformer Informer PatchTST SAITS ImputeFormer TimesNet SCINet DLinear FiLM

Builds on: MDLM D3PM CSDI SAITS DiffusionTS

Closest papers by meaning

• Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification (NeurIPS 2025 · similarity 0.92)
• ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation (NeurIPS 2026 · similarity 0.91)
• HyFAD: Hybrid Time-Frequency Diffusion with Frequency-Aware Embedding for Time Series Imputation (NeurIPS 2026 · similarity 0.89)
• Rethinking the Diffusion Models for Missing Data Imputation: A Gradient Flow Perspective (NeurIPS 2024 · similarity 0.89)
• LSCD: Lomb--Scargle Conditioned Diffusion for Time series Imputation (ICML 2025 · similarity 0.88)
• Frequency-aware Generative Models for Multivariate Time Series Imputation (NeurIPS 2024 · similarity 0.88)