Real-world sensor networks—whether tracking urban traffic flow, monitoring air quality, or mapping power grid loads—are inherently messy. Hardware malfunctions, battery depletion, and communication dropouts leave behind significant gaps in the data. Crucially, these gaps are rarely random, isolated points. Instead, they manifest as structured "blackouts": an entire sensor node might go offline for a day, or a localized network cluster might fail simultaneously due to a power outage.
When large contiguous blocks of space and time lack any observations, standard interpolation methods like mean substitution or linear smoothing fail completely. To reconstruct what happened during these blind spots, modern research leverages a fundamental intuition: spatiotemporal data is highly redundant. A missing sensor reading is tightly constrained by its neighbors in space (the sensors nearby) and its context in time (what happened moments before, and what happens routinely at this hour).
To capture these dual constraints, researchers have converged on three core paradigms: low-rank tensor completion, graph neural networks, and non-autoregressive attention mechanisms.
The Global View: Low-Rank Tensor Completion
One way to understand a spatiotemporal dataset is to organize it into a multi-dimensional array, or a tensor. For instance, a traffic dataset can be structured as a 3D tensor with dimensions: $\text{Sensor} \times \text{Time of Day} \times \text{Day}$.
Because human activity and environmental phenomena follow highly regular, cyclical patterns, this tensor is mathematically low-rank. This means that the massive array of numbers can actually be compressed into a combination of a few underlying factors. If a sensor goes dark for an afternoon, we can infer its missing values by aligning the remaining data with global trends across other days and other sensors.
To exploit this, Xinyu Chen et al. introduced the Low-Rank Autoregressive Tensor Completion (LATC) framework in Low-Rank Autoregressive Tensor Completion for Spatiotemporal Traffic Data Imputation. The core insight of LATC is that while low-rank factorization captures global, macro-level structures (like the morning rush hour trend), it can miss local, micro-level fluctuations (like a sudden traffic jam). By combining global tensor completion with a local autoregressive model, the framework constrains the imputed values to be both globally plausible and temporally continuous.
Global Tensor Structure (Macro Trends) + Local Autoregressive Model (Micro Continuity)
\ /
\ /
--> Robust Spatiotemporal Imputation
As datasets scale to millions of observations, traditional tensor factorization becomes computationally prohibitive. To solve this bottleneck, Yixian Chen et al. proposed LSTC-Tubal in Scalable low-rank tensor learning for spatiotemporal traffic data imputation. By utilizing data-driven unitary transformations alongside temporal smoothing, their model accelerates the imputation process by up to 40 times while preserving structural integrity. Further refining this algebraic approach, Tong Nie et al. deployed a Truncated Tensor Schatten $p$-norm in Truncated tensor Schatten p-norm based approach for spatiotemporal traffic data imputation with complicated missing patterns, demonstrating that non-convex regularizers can better approximate the true rank of a corrupted sensor network even when up to 90% of the data is missing.
The Topological View: Graph Neural Networks
While tensors excel at capturing grid-like, periodic structures, they struggle with complex physical topologies. A road network or a power grid is not a uniform grid; it is a graph where connections depend on physical infrastructure.
Graph Neural Networks (GNNs) handle missing data by executing message passing across these topological connections. If Sensor A goes offline, its immediate neighbors on the graph (Sensors B and C) pass their observed states across the graph edges to help reconstruct Sensor A's state.
In Filling the G_ap_s: Multivariate Time Series Imputation by Graph Neural Networks, Andrea Cini et al. introduced the Graph Recurrent Imputation Network (GRIN). GRIN operates by interleaving bidirectional recurrent neural networks (which capture forward and backward temporal context) with graph convolutions (which capture spatial context). Instead of treating imputation as a static, isolated task, GRIN treats the multivariate time series as a sequence of graphs, allowing spatial and temporal representations to update each other dynamically at every time step.
[Time t-1] ---> [Time t (Missing Node)] ---> [Time t+1]
^
| (Spatial Message Passing via GNN)
[Neighbor Nodes]
To ensure that these reconstructed graph signals remain physically realistic, Jhon A. Castro-Correa et al. developed TimeGNN in Time-Varying Signals Recovery Via Graph Neural Networks. TimeGNN optimizes an encoder-decoder architecture using a specialized Sobolev loss function. This mathematical constraint penalizes "roughness" across both the spatial graph and the timeline, preventing the model from generating erratic, jagged spikes in the missing segments and forcing the reconstructed signal to evolve smoothly.
Bypassing Autoregressive Drift with Attention
Many early spatiotemporal imputation models relied on recurrent architectures that filled in missing values step-by-step. However, this sequential approach suffers from a fatal flaw: error accumulation. If the model makes a slight mistake when imputing the data at 12:00 PM, that error is fed back into the model as ground truth to predict 12:05 PM, causing the predictions to rapidly "drift" away from reality over long time horizons.
To break this dependency, researchers turned to non-autoregressive attention mechanisms. Instead of stepping through time sequentially, attention allows a model to look globally at all surviving data points simultaneously.
In Learning to Reconstruct Missing Data from Spatiotemporal Graphs with Sparse Observations, Cesare Alippi et al. proposed the Spatiotemporal Point Inference Network (SPIN). SPIN completely abandons recursive prediction. Instead, it employs a sparse attention mechanism that computes a direct mapping from the sparse set of actually observed values across the entire network to the target missing points. Because every missing value is reconstructed using only true observations rather than previously imputed guesses, the model is immune to autoregressive drift, making it highly resilient against prolonged sensor blackouts.
Turning Missingness into a Feature
In many practical scenarios, the ultimate goal is not just to fill in the historical record, but to forecast future trends despite the missing data. Traditional pipelines treat imputation and forecasting as two separate steps: first they fill the holes, then they feed the completed data into a forecasting model. However, this separation ignores a vital piece of information: the pattern of missingness itself can be highly informative.
For example, a traffic sensor might fail more frequently during peak hours due to network congestion, or an environmental sensor might drop frames during extreme weather events. The fact that data is missing is a signal in its own right.
To exploit this, Xiaodan Chen et al. developed BiTGraph in Biased Temporal Convolution Graph Network for Time Series Forecasting with Missing Values. Instead of trying to erase the gaps, BiTGraph explicitly injects a "missingness bias" matrix directly into its temporal convolutions and graph neural networks. By modeling the joint distribution of when and where data is missing alongside the available context, the model turns a data quality flaw into a predictive feature.
Similarly, Chengqing Yu et al. introduced GinAR in GinAR: An End-To-End Multivariate Time Series Forecasting Model Suitable for Variable Missing, integrating interpolation attention and adaptive graph convolutions into a unified recursive architecture. This allows the system to generate highly accurate future forecasts even when up to 90% of the historical input variables are entirely unavailable, proving that end-to-end models can maintain structural awareness without requiring pristine data integrity.