Learning from Incomplete Multivariate Time Series

74 concepts71 papers1 proposal

Real-world multivariate time series (MTS) data—whether recorded from clinical sensors in an intensive care unit, industrial machinery, or environmental monitors—rarely arrives in clean, unbroken grids. Sensors fail, networks drop packets, and clinical observations occur at irregular intervals. This leaves behind a landscape of missing values and misaligned timestamps.

Standard machine learning models expect complete, regularly spaced matrices. The naive response to this problem is ad-hoc preprocessing: filling gaps with zeros, forward-filling the last known value, or substituting the global mean. However, as noted by Filippo Maria Bianchi et al. in Learning compressed representations of blood samples time series with missing data, these heuristic approaches destroy the underlying temporal structure and erase the informative patterns of why data is missing in the first place. For instance, in healthcare, the absence of a lab test is itself a signal that a patient may be stable.

To accurately reconstruct or learn from incomplete data, modern architectures must exploit two primary axes of information: temporal dependencies (how a variable changes over time) and inter-variable correlations (how different sensors or channels relate to one another).

Bidirectional and Disentangled Reconstruction

Early deep learning approaches tried to impute values sequentially. However, unidirectional models only see the past, leaving them ill-equipped to handle sudden gaps or capture the context of a missing block. To capture the full context of a missing interval, models must look both forward and backward.

In Bidirectional Recurrent Imputation for Time Series, Wei Cao et al. introduced BRITS, a framework that treats missing values as learnable variables embedded directly within a bidirectional Recurrent Neural Network (RNN). By updating hidden states simultaneously from both temporal directions, the network enforces consistency between past observations and future trajectories, allowing the model to infer missing points based on the entire sequence context.

Instead of relying on complex, entangled deep states, another strategy is to decompose the time series into fundamental structural components. SHUAI LIU et al. proposed TIDER (Multivariate Time-series Imputation with Disentangled Temporal Representations), which uses matrix factorization to disentangle the data into global trends, seasonal patterns, and local biases. By solving the imputation problem within this low-rank, disentangled space, the model reconstructs long-term gaps with high scalability and interpretability, avoiding the error accumulation common in recurrent step-by-step generation.

Generative Imputation via Diffusion Models

While recurrent and factorization methods provide deterministic estimates, real-world missingness is often stochastic and multi-modal; there may be multiple plausible values that could fill a gap. Recent breakthroughs leverage generative diffusion models to sample realistic completions from the conditional distribution of the missing data given the observed data.

For example, Mohammad Shohidul Islam et al. developed SADI (Self-attention-based Diffusion Model for Time-series Imputation), which pairs feature-wise and temporal-wise self-attention mechanisms within a diffusion framework. This allows the model to capture both cross-channel correlations and long-range temporal dynamics during the denoising process.

Similarly, Shunyang Zhang et al. introduced Score-CDM (Score-Weighted Convolutional Diffusion Model for Multivariate Time Series Imputation), utilizing a score-weighted convolution module to capture global patterns alongside a frequency-domain adaptive reception module for local details. This ensures that the generated values respect both smooth, long-term trends and sharp, localized variations.

Imputation-Free Representation Learning

Explicitly filling in missing values (imputation) is not always necessary, and can sometimes introduce artificial noise or systematic bias into a dataset. A parallel paradigm focuses on learning robust representations directly from incomplete, irregularly sampled streams without an explicit data-completion step.

One way to achieve this is by dynamically modeling the changing relationships between variables as data arrives. In Adapting to the stream: an instance-attention GNN method for irregular multivariate time series data, Kun Han et al. proposed DynIMTS, a dynamic Graph Neural Network (GNN). Instead of relying on a fixed sensor topology, DynIMTS uses an instance-attention mechanism to compute and update sensor correlations on the fly. This allows the model to naturally align irregular timestamps and handle missing observations during downstream classification or forecasting tasks.

Alternatively, mathematical kernels can be designed to natively tolerate gaps. Lorenzo Livi demonstrated in Time series kernel similarities for predicting Paroxysmal Atrial Fibrillation from ECGs that specialized kernel similarity measures can evaluate the proximity of two time series without requiring any imputation. By bypassing the imputation step entirely, this approach eliminates artificial bias, proving highly effective for noisy, intermittent data streams from wearable medical devices where signal gaps are frequent due to movement or poor sensor contact.

Key Foundational Work

Detection and Characterization of Anomalies in Multivariate Time Series (2009) — contributed 1 concept elements to this direction.
Learning Nonlinear Dynamical Systems Using an EM Algorithm (1998) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_missing_data_in_time_series_8385b7ad,
  title = {missing data in time series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/missing-data-in-time-series/rdr_8385b7adc4264dca90caf7a355dc0b35},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_missing_data_in_time_series_8385b7ad_v1,
  title = {missing data in time series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/missing-data-in-time-series/rdr_8385b7adc4264dca90caf7a355dc0b35/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake