Every real sensor network leaves parts of the system unmeasured. Some road segments have no loop detector, some substations have no meter, some neurons are never recorded. Spatio-temporal graph neural networks are the standard forecasters for such data. They mix information across a graph of sensors at each time step and model each sensor's history. When the true graph is unknown, they learn one, typically as a product of per-sensor embedding vectors. An unmeasured unit that drives two visible sensors makes those sensors correlated, so the learned graph draws a direct "ghost" edge between them. The forecast may still look fine, but the graph is wrong. It cannot tell you where influence actually flows, and it carries no explicit representation of the missing unit.
We propose to give the forecaster a small number of latent nodes: extra graph nodes that have their own state evolving over time, receive and send messages to visible sensors through sparse learned edges, and are never directly supervised. Visible-to-visible edges are pushed to be sparse and to explain only direct coupling. Anything that looks like a shared, delayed common cause must be routed through the latent nodes. The design comes from a simple algebraic fact in partial-observability theory, stated below. Eliminating hidden units from a linear network leaves a correction that is low-rank across sensors and has memory in time. Adaptive graphs in current forecasters capture the low-rank part but have no memory channel. We test whether this structure reduces ghost edges and makes hidden-sensor signals recoverable, at equal forecasting accuracy, when we deliberately hide real sensors.
Core Claim
Consider linear dynamics on visible states $x$ and hidden states $h$. Let $A_{vv}, A_{vh}, A_{hv}, A_{hh}$ denote the visible-to-visible, hidden-to-visible, visible-to-hidden and hidden-to-hidden coupling blocks, so that $x_{t+1} = A_{vv}x_t + A_{vh}h_t$ and $h_{t+1} = A_{hh}h_t + A_{hv}x_t$. Unrolling $h$ gives
$$x_{t+1} = A_{vv}x_t + \sum_{k\ge 0} A_{vh}A_{hh}^{k}A_{hv}\,x_{t-k-1} + \text{noise}.$$
Every term in the sum has rank at most the number of hidden units, and the sum reaches back in time. This is the "structured low-rank perturbation or rational transfer function" that subspace methods exploit in State-Space Network Topology Identification From Partial Observations, which identifies transition matrices from Hankel matrices and singular value decomposition, up to a change of basis. Graph WaveNet's adaptive adjacency $\mathrm{Softmax}(\mathrm{ReLU}(E_1E_2^\top))$ (Graph WaveNet) is already low-rank, but it is memoryless and is not separated from the direct links. The formula therefore predicts that it will absorb hidden-node influence as spurious direct edges.
Our claim is the following. A forecaster that splits spatial mixing into a sparse direct graph and a small set of latent nodes with their own recurrent state will match the forecasting error of adaptive-graph models on the visible sensors. It will also substantially reduce ghost edges, and its latent states will linearly predict the readings of sensors that were hidden during training. The most stringent of these three predictions is the memory requirement: a memoryless low-rank correction should not suffice.
Research Program
The backbone is a decoupled spatio-temporal network in the style of Graph WaveNet. It uses dilated temporal convolutions per sensor interleaved with graph diffusion. We replace its adaptive graph with two pathways.
The direct pathway is a sparse matrix $S$ over visible sensors. It is parameterized from node embeddings, but it carries an L1 penalty, and its edges are made hard after training by the k-means separation rule from Graph Learning Under Partial Observability. That paper shows consistent graph recovery for vector autoregressive models needs only an "identifiability gap" between edge and non-edge scores, not unbiased edge estimates. We reuse its estimator logic as the method for reading off structure.
The latent pathway is a set of $K$ latent nodes. Each has a gated recurrent state $h_t$, and the latent nodes are connected to visible sensors through sparse read-in matrix $B$ and read-out matrix $C$. This pathway contributes $C\,h_t$ to the next visible update and has no instantaneous visible-to-visible term. The rank bound and the delay are therefore built into the architecture rather than hoped for. $K$ is chosen on validation forecasting error, and we report sensitivity to it.
The analysis has two parts. On synthetic linear and mildly nonlinear diffusion with planted hidden nodes, we characterize when sparse direct coupling and latent coupling can be told apart. We expect failure when hidden nodes are dense clusters adjacent to many visible nodes, analogous to the regime where the four-hop guarantee of Topology Learning of Radial Dynamical Systems with Latent Nodes breaks. We also expect a residual cospectral ambiguity of the kind formalized by Coutino et al. We report these limits rather than claim universal identifiability.
Related work plays three different roles here. SICSM motivates the proposal. It shows that disambiguating latent-mediated paths ($A\to H\to B$) from direct links is possible, using multi-layer aggregation over a selective state-space encoder with graph sampling. But it only infers a static graph, does not forecast, and is the slowest method in its own evaluation. We seek the same disambiguation inside a forecaster, without combinatorial graph sampling. SIDEC constrains the design. It scales structural inference to 1,000 nodes via per-node encoding and partial correlation, but it assumes every interacting unit is observed, so its partial correlations will mark hidden-driven pairs as edges. NetFormer shows the same vulnerability to unrecorded hub neurons in neuroscience, which motivates hub-hiding as our hardest test.
The kriging work answers a different question. GRIN, GgNet and DarkFarseer reconstruct signals at unmonitored locations, but each assumes the location and its graph neighbors are known. In GgNet's case the location also has partial covariates. We assume the hidden unit's existence and position are unknown, and we ask whether the model discovers it from visible dynamics alone.
Baselines
The naive reference point is the same temporal-convolution backbone with the distance-thresholded road graph restricted to visible sensors and no learned graph. It shows how much any learned spatial structure buys once sensors are removed.
The strongest forecasting competitor is Graph WaveNet with its self-adaptive adjacency. Its embeddings give it a low-rank capacity that could in principle absorb hidden-node effects. It reports a 60-minute-horizon MAE of 3.53 on METR-LA, against 3.60 for DCRNN (Graph WaveNet). Those numbers use all 207 sensors. We found no reported result on Lacuna for METR-LA or PEMS-BAY with sensors withheld as latent, so all hidden-sensor numbers will come from our reruns. We also run SAGDFN as the scalable adaptive-graph representative. Its paper reports accuracy comparable to top models on METR-LA but gives no figure in the file.
Two-stage pipelines are a second family. Here we infer a graph with SICSM or SIDEC and then forecast with it. SICSM reports AUROC above 80% on VN_SP_30 with 60% of nodes observed, where NRI and ACD fall below 75%. SIDEC reports about 98% AUROC on NetSims under full observation. These pipelines are the strongest competitors on structure recovery in synthetic settings.
For recovering hidden-sensor signals, we compare against GRIN and DarkFarseer. DarkFarseer reports a kriging MAE of 3.796 on PEMS-BAY, and 59.344 on PEMS03 against 62.028 for IGNNK. These baselines are given the hidden sensor's location and road-graph neighbors, an advantage our model does not get.
Fairness rules apply to all comparisons. All forecasters share the backbone, the data splits and the set of hidden sensors. Each gets the same hyperparameter-trial budget and GPU-hour cap. Every adaptive-graph baseline gets an embedding dimension at least as large as our $K$ plus its own dimension, so it has equal or more low-rank capacity.
Beating Graph WaveNet means three things together. The visible-sensor MAE must be within one standard deviation across seeds. The ghost-edge rate must be at least halved. Linear probes of the latent states must recover hidden sensors with lower error than probes of Graph WaveNet's hidden features.
Three ablations isolate the key ingredient. The first replaces the latent recurrent state with a memoryless rank-$K$ term $CB$ applied to $x_t$. The second removes the sparsity penalty on $S$. The third sets $K=0$.
Why not just…?
Why not just increase the embedding dimension of Graph WaveNet's adaptive graph? This is the most serious alternative, because a larger $E_1E_2^\top$ is already a low-rank spatial correction. The unrolled formula shows hidden influence arrives through lagged visible states. A memoryless matrix can only approximate that by placing weight on the edges whose endpoints share a hidden neighbor, which is exactly what produces ghost edges. The capacity-matched comparison settles this directly. If a wide Graph WaveNet matches our ghost-edge rate and probe recovery, the claim fails.
Why not just add a few global "virtual" nodes, or a memoryless sparse-plus-low-rank split? This is the first ablation. If it performs as well as recurrent latent nodes, then the memory channel is unnecessary. We would then downgrade the claim to "sparse plus low-rank is enough", which would be cheaper but less interesting. We expect the gap to grow under hub hiding and longer forecasting horizons, where the lag sum matters most.
Why not infer the graph first with SICSM, SIDEC or subspace identification, then forecast on it? SIDEC assumes full observability by construction. SICSM is reported as the slowest baseline and outputs only a static graph. Subspace identification assumes linear time-invariant dynamics and becomes expensive with hundreds of sensors at five-minute resolution. On synthetic data with known truth, these methods may genuinely beat us on AUROC. We will report it if they do. Our case rests on matching their structure while also forecasting, at lower cost.
Proposed Experiments
On synthetic data, we simulate linear and tanh-nonlinear diffusion on tree-structured, cyclic and StructInfer vascular-network topologies. We hide 10–40% of nodes either at random or by highest degree. We then measure direct-edge AUROC on visible nodes, the ghost-edge rate, wall-clock time against SICSM, and probe recovery of the hidden trajectories.
On METR-LA and PEMS-BAY, we remove chosen sensors entirely from training and test inputs. We then measure four things. The first is MAE, RMSE and MAPE at 15, 30 and 60 minutes on the visible sensors. The second is the ghost-edge rate, defined as the fraction of learned visible edges between sensors that are not adjacent in the road graph but share a hidden neighbor. The road graph is a proxy for physical truth, and we say so. The third is the error of a linear probe from latent states to the hidden sensor series. Its readout is fit on one day of hidden data and evaluated on the held-out test period, and each baseline's internal states get the same probe. The fourth is how these metrics scale with the number of hidden sensors.
The decisive check is the combination of four outcomes. Visible MAE must be on par with Graph WaveNet. Ghost edges must fall sharply under hub hiding. Probe recovery of hidden sensors must be clearly better. The memoryless ablation must lose on the last two. If accuracy is equal but the memoryless ablation matches us, the latent-dynamics hypothesis is refuted, and we will report it that way.
References
- State-Space Network Topology Identification From Partial Observations
- Graph WaveNet for Deep Spatial-Temporal Graph Modeling
- Graph Learning Under Partial Observability
- Topology Learning of Radial Dynamical Systems with Latent Nodes
- Structural Inference of Dynamical Systems with Conjoined State Space Models
- Structural Inference with Dynamics Encoding and Partial Correlation Coefficients
- NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics
- Filling the G_ap_s: Multivariate Time Series Imputation by Graph Neural Networks
- Graph-based Virtual Sensing from Sparse and Partial Multivariate Observations
- DarkFarseer: Robust Spatio-Temporal Kriging Under Graph Sparsity and Noise
- SAGDFN: A Scalable Adaptive Graph Diffusion Forecasting Network for Multivariate Time Series Forecasting