Deep Learning for Irregularly Sampled Time Series

1,482 concepts1,083 papers11 proposals

Standard deep learning architectures share a fundamental assumption: data arrives on a perfectly uniform grid. In an audio file, samples occur at a fixed frequency. In text, words follow one another in discrete, sequential steps. Consequently, standard Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Transformers are hardcoded to treat the transition from step $t-1$ to step $t$ identically, implicitly assuming that the elapsed time between observations ($\Delta t$) is constant.

In the real world, this assumption frequently breaks down. Consider Electronic Health Records (EHRs), where a patient might visit a clinic twice in one week due to an acute illness, and then not return for two years. As noted by Edward Choi et al. in their work on Doctor AI, the irregular duration between these visits is itself a critical clinical signal. Similarly, in remote sensing, cloud cover sporadically blocks satellite observations, leaving unpredictable gaps in what should be a steady temporal stream (Vivien Sainte Fare Garnot et al.).

As surveyed by Bryan Lim et al. in Time Series Forecasting With Deep Learning: A Survey, forcing irregular data into fixed-interval grids via heuristic imputation or binning introduces severe biases, masks underlying uncertainty, or destroys the temporal rhythm of the data. To address this, a distinct research direction has emerged to build architectures that natively understand continuous, non-uniform time intervals.

Time-Aware Recurrence and Gated Decay

The earliest and most intuitive way to adapt deep learning for irregular intervals is to modify the internal gating mechanisms of recurrent networks. If a patient exhibited a symptom three days ago, it should heavily influence a diagnostic prediction today; if that same symptom occurred three years ago, its relevance has likely faded.

To capture this intuition, Trang Pham et al. developed DeepCare, an architecture that introduces time-dependent decay mechanisms directly into the hidden states of a Long Short-Term Memory (LSTM) network. Instead of passing the hidden state $h_{t-1}$ directly to the next step, the historical context is discounted as a function of the elapsed time $\Delta t$.

Crucially, this decay shouldn't be uniform across all variables. A chronic condition like diabetes persists over decades, whereas an acute fever resolves in days. To handle this multi-scale reality, models like the one proposed by Yonghong Luo et al. in Multivariate Time Series Imputation with Generative Adversarial Networks learn parameterized, feature-specific decay rates directly from the data. This allows the network to dynamically decide how fast information fades for different variables. In a single-author study on visual analytics for EHRs, Jimeng Sun demonstrated in RetainVis that explicitly incorporating these irregular temporal gaps is essential for preserving diagnostic meaning over long, sporadic patient trajectories.

Continuous Latent Flows via Neural Differential Equations

While gated decay mechanisms improve recurrent models, they still treat observations as discrete steps. They fail to explicitly model what happens between observations. To solve this, researchers shifted toward a continuous-time paradigm: treating the hidden state not as a sequence of discrete updates, but as a continuous trajectory governed by a differential equation.

In this framework, between observations $t_{i-1}$ and $t_i$, the hidden state $h(t)$ evolves continuously according to a Neural Ordinary Differential Equation (Neural ODE):

$$\frac{dh(t)}{dt} = f(h(t), t; \theta)$$

When a new observation arrives at $t_i$, the model ingests the new data and updates the hidden state abruptly. This combination of continuous evolution and discrete updates forms a "jump" architecture. Calypso Herrera et al. proved in Neural Jump Ordinary Differential Equations: Consistent Continuous-Time Prediction and Filtering that this formulation mathematically converges to the optimal conditional expectation for irregularly sampled data.

To account for sudden, unpredictable real-world shocks—such as an abrupt medical intervention or an earthquake—Junteng Jia et al. extended this concept to stochastic systems in Neural Jump Stochastic Differential Equations. By integrating discrete jumps into continuous latent flows, these models naturally handle arbitrary, real-valued time intervals without requiring any data imputation.

Continuous Time Embeddings and Attention

Recurrent and differential models are inherently sequential, which limits their computational efficiency. To leverage the parallel processing power of attention mechanisms on irregular data, researchers had to rethink how transformers represent time.

In a standard Transformer, positional encodings rely on discrete integer indices ($1, 2, 3, \dots$) to represent sequence order. For irregular time series, this is replaced by mapping real-valued timestamps into a continuous vector space. By using continuous time embeddings, the model can compute attention weights based on the actual physical duration between events rather than their indexical order.

Discrete Encodings:   [Step 1] --------> [Step 2] --------> [Step 3]  (Assumes equal gaps)

Continuous Embeddings: [t = 1.2] -> [t = 1.5] --------------------> [t = 24.8] (Preserves true scale)

Satya Narayan Shukla et al. utilized this insight to design Multi-Time Attention Networks for Irregularly Sampled Time Series (mTANs). Instead of stepping sequentially through time, mTANs use a continuous-time embedding layer that allows the model to query the irregular time series at any arbitrary set of reference time points. This enables the network to learn global temporal dependencies across non-uniform intervals while training orders of magnitude faster than continuous differential equation models.

Decoupling via Learnable Interpolation

A final, highly practical strategy focuses on decoupling the irregularity of data collection from the downstream prediction task. Real-world multivariate time series are often not just irregular, but also unaligned—meaning different sensors or biomarkers are measured at completely different times, creating a chaotic, asynchronous matrix.

Instead of forcing the predictive model to handle this unaligned raw data directly, researchers introduced an end-to-end learnable interpolation layer. In Interpolation-Prediction Networks for Irregularly Sampled Time Series, Satya Narayan Shukla et al. demonstrated that a model can automatically project messy, unaligned observations onto a clean, uniform virtual grid.

This interpolation layer decomposes the raw data into distinct channels—such as smooth trends and transient spikes—while simultaneously tracking observation intensity (how frequently a variable is measured). Because the interpolation and downstream prediction networks are trained jointly, the system avoids the hand-crafted biases of traditional imputation methods, allowing standard, highly optimized deep learning architectures to run safely on the regularized virtual grid.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_irregular_temporal_sampling_6f62c1dc,
  title = {irregular temporal sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/irregular-temporal-sampling/rdr_6f62c1dc673d4e7699612fbaf233ccc3},
  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_irregular_temporal_sampling_6f62c1dc_v1,
  title = {irregular temporal sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/irregular-temporal-sampling/rdr_6f62c1dc673d4e7699612fbaf233ccc3/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