Research map / Time series / Causal and generative time series

ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation

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 ProCTI, a diffusion time-series imputation model that adds learnable prototype-bank global context (cross-attention retrieval) to FGTI-style frequency conditioning, with a latent-regime theory on when global conditioning helps.

Problem

Diffusion-based time series imputation conditions mainly on local context, leaving dataset-level global structure implicit, which hurts performance when local observations are sparse, noisy, or unrepresentative.

Approach

A learnable prototype bank queried by each partially observed window via cross-attention yields an attribute-wise regime vector that is added to the dominant-frequency conditioning branch of a frequency-aware diffusion imputer; theory via a latent-regime model characterises when this global conditioning reduces conditional entropy and denoising MSE.

Key results

• Under random missingness, ProCTI achieves the best mean MAE on all five datasets, e.g. Stock MAE 2.223 vs FGTI 3.186 and PhysioNet MAE 0.259 vs 0.362 (aggregated over missing ratios).
• The paper reports ProCTI reduces average MAE by roughly 3-30% relative to FGTI, the strongest diffusion baseline.
• Under attribute-drop missingness ProCTI is best on Gait, PhysioNet, and Beijing, while BRITS is best on Weather and Stock where cross-feature redundancy is strong.

Surprising findings

• BRITS beats ProCTI under attribute-wise missingness on Weather and Stock, which the authors attribute to strong cross-feature redundancy making global conditioning unnecessary.
• Frozen prototype bank (variant D) performs worse than the variant with the learnable bank, and injecting regime information into the high-frequency branch hurts.

Limitations

• Gains over FGTI are small on some datasets (e.g. Gait, Beijing) and several significance tests with only 5 seeds are not significant.
• Theory relies on a latent-regime model and approximations (e.g. attention approximating p(Z|O)) that are not verified directly.
• Higher GPU memory than most baselines, similar to FGTI.

Experimental setup

Datasets: PhysioNet Beijing Air Quality Weather Stock Gait

Benchmarks: PhysioNet Beijing Air Quality Weather Stock Gait

Metrics: MAE RMSE CRPS PI90 coverage PI90 width

Baselines: BRITS CSDI SCINet TIDER MTSCI Diffusion-TS iTransformer FGTI PaD-TS

Builds on: FGTI DDPM CSDI Diffusion-TS PaD-TS

Closest papers by meaning

• Discretizing Continuous Time Series for Imputation with Masked Diffusion Training (NeurIPS 2026 · similarity 0.91)
• Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification (NeurIPS 2025 · similarity 0.91)
• Rethinking the Diffusion Models for Missing Data Imputation: A Gradient Flow Perspective (NeurIPS 2024 · similarity 0.89)
• Glocal Information Bottleneck for Time Series Imputation (NeurIPS 2025 · similarity 0.88)
• LSCD: Lomb--Scargle Conditioned Diffusion for Time series Imputation (ICML 2025 · similarity 0.88)
• HyFAD: Hybrid Time-Frequency Diffusion with Frequency-Aware Embedding for Time Series Imputation (NeurIPS 2026 · similarity 0.87)