Time series data, sequences of measurements ordered in time, underpin countless real-world applications, including clinical diagnostics, air quality monitoring, electricity load forecasting, anomaly detection, and emergency event prediction. Because missing values are both common and inevitable, the quality of any downstream analysis depends heavily on how well those gaps can be filled. Traditional approaches relied on statistical techniques such as ARIMA or simple substitution with zeros or means, while later machine learning models like BRITS and GRU-D brought recurrent neural networks to the task. More recently, Transformer-based architectures such as SAITS and ImputeFormer have set strong benchmarks. Yet all of these methods share a common limitation: they treat imputation as a deterministic estimation problem, producing a single best guess while ignoring the genuine uncertainty that surrounds any missing measurement.

Diffusion models, the same generative technology behind today’s most striking image generators, offer a natural alternative. They work by gradually adding Gaussian noise to data in a forward process and then learning to reverse that process, step by step, transforming pure noise into realistic samples. Because they model entire probability distributions rather than point estimates, diffusion models can capture the uncertainty of missing values and even provide confidence bounds around their predictions. Models like CSDI, SSSD, and SPD have already applied this idea to time series imputation with promising results. But according to the Tokyo team, these existing approaches suffer from two fundamental weaknesses that undermine the consistency of their output.

The first weakness is what the researchers call padding-induced distribution shift. To train a neural network on incomplete data, missing entries must be filled with placeholder values, typically zeros, so that the data can flow through the network’s mathematical machinery. The trouble is that diffusion models learn to approximate the distribution of whatever they are fed during training. When a large fraction of the training inputs consists of artificial zeros rather than genuine measurements, the model’s learned target distribution drifts away from the true data distribution. The effect is far from trivial: the team’s exploratory analysis showed that simply swapping zero-padding for one-padding degraded performance by 54.7 percent on one benchmark, while switching from sample-mean padding to linear interpolation improved results by 62.8 percent. In other words, the seemingly mundane choice of how to fill in blanks before training can swing accuracy dramatically.

The second weakness lies in how diffusion models are optimized. Standard diffusion training is entirely noise-driven: the network learns to predict the Gaussian noise that was added at each step, following the distributional dynamics of the reverse process. But this scheme provides no direct supervision from the actual observed values in the data. The result, the researchers show, is an unstable denoising trajectory, the sequence of intermediate states the model passes through as it walks from pure noise back to a clean sample. Visualizations of the popular CSDI model reveal oscillatory, erratic trajectories with large errors persisting deep into the reverse process, which ultimately produces imputations that can be inconsistent with the observed data surrounding the gaps.

ITCD tackles both problems head-on with a two-part design. The first component is a two-stage diffusion architecture. In the Refinement Stage, the model trains on padding-corrupted samples but uses its own probabilistic generations to replace the artificial padding with distribution-consistent values. These refined samples then serve as training inputs for the Core Stage, a second diffusion model that learns on data far closer to the true underlying distribution. Because the Core Stage is the only component used at inference time, the Refinement Stage adds no computational cost when the trained model is deployed. Ablation experiments confirm the value of this design: adding the Refinement Stage reduced error by 18.6 percent on the PeMS traffic dataset and 65.4 percent on the Electricity dataset in one comparison, and by 14.3 percent on Beijing Air quality data in another.

The second component, Implicit Trajectory-Constrained Optimization, reshapes how the denoising process itself is guided. It introduces two complementary constraints. The Intermediate Constraint supervises the predicted noise at every step of the reverse process, and crucially, it does so across all positions in the data, not just the missing ones. Existing methods typically focus only on the missing regions, discarding the rich distributional information contained in the observed values. By minimizing the mean absolute error between predicted and added noise everywhere, the model is pushed to align its entire denoising trajectory with the true data distribution. The Terminal Constraint then supervises the fully reconstructed sample at the final reverse step, combining an Observed Reconstruction Task that learns directly from visible measurements with a Masked Imputation Task that forces accurate prediction of artificially hidden values. Together, these constraints implicitly steer the model along a coherent path from noise to data, without requiring any extra learnable parameters.

The architecture also includes a clever structural safeguard during sampling. Before each denoising step, the entries corresponding to observed positions are explicitly refreshed with the true conditional values, preventing inaccurate intermediate estimates from contaminating the missing regions and reducing error accumulation along the trajectory. A small amount of stochastic noise is deliberately retained at each step to keep the trajectory from collapsing into a rigid deterministic path, which strengthens the robustness of the imposed constraints. The denoising network itself uses attention mechanisms in two directions simultaneously, across time and across features, with a learned gating mechanism that blends the two perspectives and incorporates information about the current diffusion step.

The experimental evidence is striking. Tested on eight datasets spanning air quality, traffic, electricity, and healthcare, including Beijing Air, Italy Air, PeMS, Melbourne Pedestrian, ETTh1, Electricity, and two PhysioNet clinical benchmarks, ITCD outperformed a broad field of baselines built on Transformers, recurrent networks, convolutional networks, and generative models. Under 10 percent point missingness, it cut error by 35.6 percent on Italy Air, 22.7 percent on PeMS, and 12.5 percent on ETTh1 relative to the best competitors. Its advantage grew as missingness increased: while CSDI’s error on Beijing Air ballooned from 0.144 to 0.640 as missing ratios rose from low to 90 percent, ITCD’s curve stayed remarkably flat. The method also proved robust across point, subsequence, and block missing patterns, and across repeated random sampling runs it showed substantially lower variance and better distributional alignment with ground truth than CSDI, as measured by the Wasserstein distance.

Perhaps most surprising for a diffusion model, ITCD is fast. Because the trajectory constraints stabilize the reverse process, the model needs only about ten diffusion steps rather than the fifty typically used by competitors, and it still matches or exceeds their accuracy. That translates into an 89.1 percent reduction in inference time on the ETT dataset and a 52.8-fold acceleration on PhysioNet2012, and it even beats the recurrent BRITS model on most benchmarks. As a probabilistic model, ITCD also provides meaningful uncertainty bounds around its predictions, giving practitioners a more honest picture of what remains unknown. The researchers, whose work was supported by the Japan Science and Technology Agency, have released their code publicly and plan to explore how imputation quality affects downstream tasks and to investigate explicit trajectory search strategies. For a field drowning in incomplete data, a method that is simultaneously more accurate, more stable, and dramatically faster is likely to attract attention well beyond the machine learning community.

Subject of Research: Diffusion-based deep learning for consistent imputation of missing values in multivariate time series

Article Title: ITCD: implicit trajectory-constrained diffusion for consistent time series imputation

Article References: Ma, S., Xu, Z., & Ichise, R. (2026). ITCD: implicit trajectory-constrained diffusion for consistent time series imputation. Applied Intelligence, 56(14), Article 397. https://doi.org/10.1007/s10489-026-07427-3

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07427-3

Keywords: time series imputation, diffusion models, missing data, deep learning, denoising trajectory, generative models, machine learning, uncertainty quantification, neural networks, Applied Intelligence, conditional diffusion, data science

Researchers led by Agung Bella Putra Utama, Aji Prasetya Wibawa, and Anik Nur Handayani of Universitas Negeri Malang, together with Andrew Nafalski of Adelaide University, have introduced a deterministic duel-based imputation framework that puts forecasting performance itself in charge of deciding how missing values should be filled. The central insight is deceptively simple: most imputation methods are judged by how statistically similar their reconstructions are to the original data, not by whether they actually help a forecasting model make better predictions. A filled-in series can look numerically plausible on paper and still wreck a downstream forecast by smoothing away the very peaks and rhythms the model needs to learn.

The framework works like a structured tournament. For every missing observation, the system generates a set of candidate values using familiar baseline techniques, including mean, median, and mode substitution, K-Nearest Neighbors, Multiple Imputation by Chained Equations, and Last-Observation-Carried-Forward. Each candidate is then scored with a composite loss function that combines three complementary forecasting metrics: Mean Absolute Percentage Error, which captures proportional error; Root Mean Square Error, which penalizes large absolute deviations; and the coefficient of determination, or R-squared, which measures how much of the variance in the real series the reconstruction explains. The weights are set at 0.4, 0.4, and 0.2 respectively, so error-based criteria dominate while explanatory power still plays a supporting role.

What separates this approach from conventional candidate ranking is the way winners are chosen. Instead of collapsing every candidate into a single aggregated score and picking the top one, the framework stages deterministic pairwise duels. Candidates face each other one-on-one, and a sensitivity threshold of 0.001 prevents negligible score differences from flipping outcomes. After all comparisons, candidates are ranked by cumulative wins, and the top three advance to a final aggregation stage. Four aggregation modes are available: averaging the survivors for low-variance stability, taking the median for outlier robustness, taking the maximum to emphasize peak responsiveness, or a winner-take-all selection of the candidate with the lowest forecasting loss. Because every step follows fixed rules, the same input always produces the same output, a property the authors argue is essential for auditing, validation, and accountability in operational systems.

The team evaluated the framework on the Beijing PM2.5 dataset, a widely used benchmark of hourly meteorological and air-pollution measurements collected between 2010 and 2014. After cleaning and temporal alignment, 41,776 valid records remained out of 43,800, with roughly 2,068 missing values concentrated in the PM2.5 variable itself. Crucially, the missingness is not random: gaps cluster in cold seasons and during periods of rapid atmospheric change, exactly the conditions where forecasting matters most and where naive imputation does the most damage. The dataset includes dew point, temperature, pressure, cumulative wind speed, snowfall, rainfall, and temporal indicators as explanatory variables, with PM2.5 concentration serving as the prediction target.

The imputed datasets were then fed into four recurrent forecasting architectures trained under identical conditions: a vanilla Recurrent Neural Network, a Long Short-Term Memory network, a Bidirectional LSTM, and a Gated Recurrent Unit. Hyperparameters for all models were tuned with Particle Swarm Optimization, though the researchers are careful to note that this stochastic tuning happens entirely outside the imputation stage and does not compromise its determinism. The optimized configurations converged on surprisingly lightweight designs: three hidden layers of 24 neurons each, sigmoid activations, the Adam optimizer with mean squared error loss, a batch size of 32, 46 training epochs, and a dropout rate of 0.2.

The results are striking. Compared with conventional imputation methods, the duel-based framework reduced MAPE by up to 70 percent and RMSE by 12 to 15 percent, while maintaining R-squared values above 0.95 across all four architectures. The Mean aggregation variant delivered the strongest overall performance, preserving both proportional variation and the amplitude structure of the signal. Dropping missing observations entirely, by contrast, produced the worst forecasts, confirming that simply discarding incomplete rows destroys the temporal coherence recurrent models rely on. Even BRITS, a sophisticated deep-learning imputation method, showed less consistent gains, suggesting that reconstruction quality alone does not guarantee forecasting quality.

Statistical testing reinforced the case. Paired t-tests and Wilcoxon signed-rank tests comparing the proposed Mean variant against KNN, the strongest conventional baseline, produced p-values below 0.01 across all architectures, with each experiment repeated ten times under fixed random seeds. Visual analysis of predicted versus observed PM2.5 trajectories showed close alignment across smooth and volatile segments alike, with the Bi-LSTM achieving the highest R-squared values and the GRU delivering the most efficient runtime while matching LSTM accuracy. The framework’s computational overhead scales quadratically with the number of candidates but linearly with data size, and because the candidate set is small and fixed, runtimes remain predictable even as data volume grows.

To test generalizability, the researchers extended the evaluation beyond air quality to three additional domains: the Heart Disease dataset, which has weak temporal structure; the KEDS e-journal dataset, which shows irregular behavioral patterns and moderate sparsity; and the Sunspot dataset, which exhibits strong periodic behavior. The framework achieved lower MAPE and RMSE than baselines including drop-missing, mean imputation, KNN, and BRITS across all of them, with the largest gains appearing in datasets with strong sequential patterns such as Sunspot. The authors acknowledge limitations, including the quadratic cost of pairwise comparison for large candidate pools and the fixed forecasting horizons tested, and they point to cluster-based candidate reduction and online learning as future directions. But the broader message stands: imputation should not be a passive preprocessing chore. Treated as a forecasting-driven decision process, it becomes a strategic lever for building AI systems that are not only accurate but consistent, transparent, and worthy of trust.

Subject of Research: A deterministic duel-based imputation framework that integrates forecasting performance metrics into missing-data selection for reliable AI-driven time-series forecasting.

Article Title: A deterministic duel-based imputation framework for reliable AI-driven time-series forecasting

Article References: Utama, A. B. P., Wibawa, A. P., Handayani, A. N., & Nafalski, A. (2026). A deterministic duel-based imputation framework for reliable AI-driven time-series forecasting. Discover Informatics, 1(1), Article 3. https://doi.org/10.1007/s44564-026-00001-6

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00001-6

Keywords: time-series imputation, missing data, AI forecasting, deterministic imputation, deep learning, LSTM, GRU, RNN, PM2.5, forecasting reliability, machine learning, data preprocessing