Spatiotemporal Imputation for Incomplete Sensor Networks

This direction is split into other directions: sensor data sparsity. Its last version stays readable and citable here.

Version 1 of 1304 papers327 conceptsfrozen 2026-10-04

In an ideal world, sensor networks—whether monitoring traffic flow in Chengdu or air quality in Delhi—would provide a continuous, unbroken stream of data. In reality, these networks are "messy." Hardware fails, communication links drop, and power outages create massive "blackouts" where entire regions go dark.

Standard machine learning models typically assume complete input grids or sequences. When faced with missing values, they often resort to crude fixes like zero-filling or mean imputation, which introduce significant bias. Modern research has shifted toward spatiotemporal imputation: a suite of techniques that treat missing data not as a nuisance to be ignored, but as a latent variable to be reconstructed by leveraging the inherent redundancy of the physical world across space and time.

The Low-Rank Intuition: Patterns Repeat

The most fundamental insight in this field is that spatiotemporal data is highly redundant. Traffic patterns on a Monday morning are remarkably similar to those on a Tuesday; the air quality at one street corner is usually tied to the quality at the next. Mathematically, this means the data can be represented as a low-rank tensor.

Xinyu Chen, Jinming Yang, and Lijun Sun exploit this in LRTC-TNN by using a "truncated nuclear norm." Instead of trying to minimize all components of the data, they focus on preserving the dominant signals (the "rank") while ignoring the noise. This allows the model to fill in large, continuous blocks of missing data by essentially "copy-pasting" the underlying structural patterns from other times and locations.

To handle more complex scenarios, Xinyu Chen and colleagues later introduced LATC, which combines this global structural view with local autoregressive dynamics. If a sensor fails, the model looks at the global "low-rank" trend to get the general shape of the data, then uses local time-series logic to fine-tune the specific values.

Virtual Sensing via Graph Neural Networks

While tensors capture grid-like patterns, many sensor networks are better described as graphs—nodes (sensors) connected by edges (roads or proximity). If a sensor at an intersection fails, its neighbors often contain enough information to reconstruct its signal. This is known as virtual sensing.

Andrea Cini, Ivan Marisca, and Cesare Alippi developed the Graph Recurrent Imputation Network (GRIN) to formalize this. GRIN passes messages between neighboring sensors both forward and backward in time. By treating the multivariate time series as a sequence of interconnected graphs, the model can "borrow" data from a functioning sensor three miles away to fill a gap in a broken one, provided they share a strong "semantic" correlation.

Similarly, Jhon A. Castro-Correa and his team introduced TimeGNN, which uses a loss function that penalizes "roughness." The intuition is that physical phenomena like temperature or traffic don't usually change instantly or sporadically; they are "smooth" across the graph. By enforcing this smoothness, the model can accurately interpolate missing values even in sparse networks.

Generative Reconstruction: From Holes to Distributions

Traditional imputation gives you a single "best guess" for a missing value. However, in high-stakes scenarios like environmental monitoring, we need to understand the uncertainty of that guess. This has led to the rise of generative models for imputation.

  • GANs: Yonghong Luo and others pioneered the use of Generative Adversarial Networks for time series. By using a specialized "GRUI" cell that accounts for irregular time lags, the model learns to generate replacements for missing data that are indistinguishable from real sensor readings.
  • Diffusion: Senzhang Wang and Shunyang Zhang developed SaSDim, a diffusion-based model. Diffusion models work by gradually removing noise to reveal a clean signal. SaSDim adapts this process specifically for spatial data, preventing random noise from washing out subtle environmental signals.
  • Schrödinger Bridges: For cases of extreme sparsity, Yu Chen and Wei Deng proposed CSBI. This framework optimizes the "transport cost" between data distributions, effectively finding the most mathematically efficient way to bridge the gap between two known data points across a void of missing values.

Modeling Continuity in a Discrete World

A final, powerful direction involves moving away from discrete time steps entirely. Ming Jiang and Yan Xu introduced GCRNODE, which uses Neural Ordinary Differential Equations (ODEs).

Instead of seeing traffic as a sequence of snapshots (where a missing snapshot is a "hole"), GCRNODE models traffic as a continuous fluid-like system. Because the underlying model is a continuous function, "missing data" simply becomes a point on a curve that hasn't been sampled yet. This allows the model to maintain high accuracy even when up to 80% of the sensor data is missing, as it is predicting the evolution of a system rather than just filling in a grid.

By combining these perspectives—global low-rank structure, local graph-based correlations, and continuous generative modeling—researchers have moved from simple "data cleaning" to creating robust, self-healing sensor networks capable of maintaining operational integrity in the face of real-world chaos.

Papers in this version

Showing 60 of 304 papers.

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_spatiotemporal_imputation_for_incomplete_dc790f19,
  title = {Spatiotemporal Imputation for Incomplete Sensor Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatiotemporal-imputation-for-incomplete-sensor-networks/rdr_dc790f19ec064f25926437af41837d4b},
  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_spatiotemporal_imputation_for_incomplete_dc790f19_v1,
  title = {Spatiotemporal Imputation for Incomplete Sensor Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatiotemporal-imputation-for-incomplete-sensor-networks/rdr_dc790f19ec064f25926437af41837d4b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}