A new machine-learning preprint reports that ProCTI produced the lowest mean absolute error, or MAE, among the methods tested when missing values were simulated across five real-world time-series datasets. Its MAE was approximately 3% to 30% lower than FGTI's.
At the center of the approach is a global context vector for each attribute. ProCTI combines it with local context during reverse diffusion, the step in which the model reconstructs missing values.
A global signal for local gaps
To test random missingness, the researchers masked observed values with a two-state Markov process. It alternated between masked and unmasked segments, with the lengths of those segments drawn from geometric distributions. The evaluation used missing ratios of 0.10, 0.30, 0.50 and 0.70. MAE and root mean squared error, or RMSE, were summarized as means and standard deviations over five random runs.
The benchmark compared ProCTI with nine baseline methods. Under the simulated random missingness, ProCTI had the best mean MAE on all five datasets.
The advantage was not universal
A separate test removed complete attributes rather than scattered stretches of a series. Drop 1 masked one attribute; Drop 2 masked two and reconstructed them from the remaining observed variables. For PhysioNet, only attributes with more than 80% of their values observed were included.
ProCTI had the lowest MAE and RMSE in most attribute-drop scenarios on Gait, PhysioNet and Beijing. BRITS, however, had the lowest MAE and RMSE on Weather and Stock under attribute-wise missingness.
The contrast makes the main result narrower than a claim of universal superiority. ProCTI led the random-missingness comparison, but its ranking changed when the missingness pattern and dataset changed.
Why the authors expect it to help
The paper's theoretical argument gives a condition for why global context might help. If a latent regime still contains information about the missing entries after the observed entries are known, then uncertainty about those missing entries is lower when the regime is also known. The result is stated for an idealized setting and optimal denoisers, so it is not a guarantee for a learned model.
An ablation study found Variant E was best in most cases. Learnable alpha injection in the dominant-frequency branch generally produced lower error, except that Variant F was lower on PhysioNet.
The evidence still has clear limits
The statistical evidence was mixed. Across five seeds, ProCTI was significant at p < 0.05 in four of eight Beijing comparisons with a t-test and one of eight with a Wilcoxon signed-rank test. On Gait, it was significant in five of eight comparisons under both tests.
In supplementary probabilistic tests, ProCTI had the best or second-best CRPS across missingness ratios, while the proportion covered by its 90% prediction interval stayed close to the nominal 90% level. These findings were descriptive, and confidence intervals were not reported.
The missing values were generated by simulation, so these tests do not establish how the model would perform under naturally occurring missingness. The analysis also compared results over only five random runs. The evidence therefore applies to this benchmark setup and its masking rules.
What the study establishes
The document is identified as arXiv:2609.37632v1 and dated 29 September 2026. Its front matter states acceptance to the 40th Conference on Neural Information Processing Systems in 2026. The paper reports funding from the Office of National Intelligence, Australia, and the National Intelligence and Security Discovery Research Grant, with identifiers GA396359 and NI240100159.
Paper data and sources
Original title: ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation
Authors: Fariza Rashid, Duc Van Le, Rahat Masood et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text