Gaussian Process Adapters for Irregular Time Series

This direction is split into other directions: multitask Gaussian process imputation. Its last version stays readable and citable here.

Version 1 of 139 papers41 conceptsfrozen 2026-10-04

Most deep learning architectures, from Convolutional Neural Networks (CNNs) to Transformers, share a common requirement: they expect data to arrive on a neat, uniform grid. In computer vision, this is a pixel array; in natural language, it is a sequence of tokens. However, real-world signals—the flickering of a distant supernova, the vital signs of a patient in the ICU, or the growth of crops viewed through patchy cloud cover—are rarely so cooperative. These signals are often "messy," characterized by irregular sampling intervals, significant gaps, and varying noise levels.

To bridge the gap between these fragmented observations and the rigid requirements of deep learning, researchers have converged on a powerful strategy: using Gaussian Processes (GPs) as a continuous, probabilistic front-end.

Modeling Functions Instead of Points

The core challenge of irregular time series is that the "missingness" is often informative or unavoidable. Traditional methods like linear interpolation or forward-filling, as noted by Margherita Rosnati and Vincent Fortuin, provide a brittle representation of the underlying signal. They treat data as a collection of discrete points, failing to capture the inherent smoothness or the uncertainty of what happens between observations.

Gaussian Process Regression (GPR) shifts the perspective from predicting specific values to modeling the distribution of possible functions that could have generated the data. As Benjamin M. Marlin and Steven Cheng-Xian Li demonstrate, a GP treats each irregularly sampled series as a continuous posterior distribution. This provides two critical benefits:

  1. Continuous Reconstruction: It allows for the estimation of the signal at any arbitrary point in time, effectively "filling in" gaps caused by sensor failure or seasonal cycles Uncovering delayed patterns in noisy and irregularly sampled time series: an astronomy application.
  2. Uncertainty Quantification: Unlike simple interpolation, a GP provides a measure of confidence. In regions with sparse data, the uncertainty grows, signaling to downstream models that the reconstructed values are less reliable Early Recognition of Sepsis with Gaussian Process Temporal Convolutional Networks and Dynamic Time Warping.

Leveraging Cross-Channel Dependencies

In many domains, we observe multiple related signals simultaneously. In astronomy, this might be the brightness of a star across different color filters; in medicine, it might be the correlation between heart rate and blood pressure.

Phillip Boyle and Marcus Frean pioneered the use of Multi-Output Gaussian Processes to capture these dependencies. By treating GPs as the result of convolving white noise with smoothing kernels, they made it possible to model shifts and lags between different data streams.

This "multi-task" approach allows a model to leverage information from a densely sampled channel to reconstruct a sparse one. For instance, Marzyeh Ghassemi and colleagues used this to create "latent fingerprints" of ICU patients, where the shared covariance structure across different physiological signals improved mortality predictions even when specific measurements were missing A Multivariate Timeseries Modeling Approach to Severity of Illness Assessment and Forecasting in ICU with Sparse, Heterogeneous Clinical Data.

The Gridification Pipeline

The most common application of this idea is to use the GP as a "gridification" layer. By fitting a GP to irregular data and then sampling from that GP at regular intervals, researchers can transform messy time series into a format compatible with standard architectures.

In astronomy, this has become a standard preprocessing step for classifying transients. Helen Qu and Masao Sako use 2D Gaussian Processes to interpolate sparse multi-band light curves into smooth, fixed-size "flux matrices." These matrices are essentially images, allowing the researchers to apply standard CNN architectures to achieve over 99% accuracy in supernova classification A Convolutional Neural Network Approach to Supernova Time-Series Classification.

Similarly, V. Ashley Villar and her collaborators have used multivariate GPs to align observations across different spectral bands into a consistent format for Variational Autoencoders (VAEs). This pipeline enables the detection of "exotic" cosmic anomalies—rare events that don't fit known patterns—by projecting the GP-reconstructed signals into a low-dimensional latent space Anomaly Detection for Multivariate Time Series of Exotic Supernovae.

Probabilistic Adapters for Deep Learning

Beyond simple preprocessing, recent work has integrated GPs directly into the neural network architecture. Joseph Futoma and Katherine A Heller developed the MGP-RNN, where a multi-output GP layer acts as a front-end for a Recurrent Neural Network. This architecture maps noisy, asynchronous clinical data onto a uniform temporal grid while explicitly accounting for the effects of medical interventions in the GP's mean function.

By combining the statistical rigor of Gaussian Processes with the expressive power of deep learning, these "hybrid" models can handle the realities of physical data—noise, gaps, and irregular timing—without sacrificing the performance gains of modern AI. Whether detecting sepsis hours earlier MGP-AttTCN: an interpretable machine learning model for the prediction of sepsis or identifying rare stars in massive sky surveys Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars, the GP serves as the essential translator between the continuous physical world and the discrete digital model.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_gaussian_process_adapters_for_irregular__061493b7,
  title = {Gaussian Process Adapters for Irregular Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/gaussian-process-adapters-for-irregular-time-series/rdr_061493b7797840169dda9f6aa343a906},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_gaussian_process_adapters_for_irregular__061493b7_v1,
  title = {Gaussian Process Adapters for Irregular Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/gaussian-process-adapters-for-irregular-time-series/rdr_061493b7797840169dda9f6aa343a906/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}