CGSTAE: Causal Graph Spatial-Temporal Autoencoder

• CGSTAE is a neural architecture that merges graph learning with causal inference to deliver interpretable and robust spatio-temporal predictions.
• It employs modules such as spatial self-attention, GCLSTM encoders, and diffusion-inspired decoders to construct dynamic causal graphs.
• Empirical evaluations show improved F1-scores and noise robustness, enabling transparent anomaly detection and effective process monitoring.

The Causal Graph Spatial-Temporal Autoencoder (CGSTAE) is a class of neural architectures designed to integrate graph-structured learning with principled, interpretable causal modeling for spatio-temporal data. CGSTAE aims to deliver both high predictive reliability and explicit structural interpretability, leveraging learned graphs, causal invariance, and temporal encoding frameworks. The concept encompasses multiple architectures, notably those for spatio-temporal forecasting (Liang et al., 2023) and for process monitoring (Zhang et al., 3 Feb 2026), which share foundational principles but implement distinct mechanisms for graph construction, causal inference, temporal modeling, and interpretability.

1. Architectural Components

CGSTAE generally consists of two or three synergistic modules:

• Graph Learning Module: Infers either correlation or causal graphs, employing techniques such as spatial self-attention (SSAM) or variational GCN encoding combined with diffusion models.
• Spatio-Temporal Sequence Model: Encodes dependencies over time and across graph nodes using units such as Graph Convolutional LSTM (GCLSTM) or variational GCN encoders.
• Decoder/Forecasting Head: Utilizes generated graphs and latent features in tasks such as sequence reconstruction, process monitoring, or spatio-temporal forecasting.

In the process monitoring context (Zhang et al., 3 Feb 2026), the SSAM is employed to produce a time-varying correlation graph $A^{(t)}$, after which an explicit causal graph $A$ is distilled using a novel three-step causal graph learning algorithm grounded in the causal invariance principle. The spatio-temporal encoder-decoder is implemented via GCLSTM units, forming a sequence-to-sequence autoencoder conditioned on the graph structure.

In the forecasting context (Liang et al., 2023), node feature sequences and (possibly partial) known adjacency matrices are embedded as Gaussian posteriors over node states using a two-layer variational GCN encoder. A diffusion-inspired decoder generates dynamic causal and transition graphs by modeling node latent trajectories via linear SDEs. Forecasts are produced by applying dynamic GCN and temporal attention layers over the constructed graphs and node embeddings.

2. Graph Structure Learning and Causality

Spatial Self-Attention and Correlation Graphs

The spatial self-attention mechanism (SSAM) computes a time-varying weighted adjacency (correlation) graph as follows. For each window $t$, a stack of inputs $X^{(t)} \in \mathbb{R}^{w \times n}$ is projected into query and key matrices:

$$Q^{(t)} = X^{(t)} W_Q, \qquad K^{(t)} = X^{(t)} W_K$$

The attention matrix (similarity graph) is

$$A^{(t)} = \sigma\left( (Q^{(t)})^\top K^{(t)} / \sqrt{w} \right)$$

where $\sigma$ is an elementwise sigmoid. The learned correlation graph evolves with data and the parameters $W_Q$, $W_K$ adapt to encode predictive dependencies.

Three-Step Causal Graph Structure Learning

To explicitly recover a robust causal graph, (Zhang et al., 3 Feb 2026) introduces a three-phase algorithm:

1. Pre-training: Train SSAM and encoder-decoder to reconstruct normal data, yielding $\{A^{(t)}\}$.
2. Causal-graph learning: Freeze previous network parameters; treat $A$ as trainable, minimizing a joint loss:
   • Mean squared error for reconstruction,
   • Invariance penalty ($L_1$ norm between $A$ and each $A^{(t)}$),
   • Prior adherence (cross-entropy to $A^\text{prior}$ if available),
   • Sparsity and discreteness regularizers.
3. Fine-tuning: Remove SSAM, use fixed $A$, further fine-tune the encoder-decoder for final performance.

This approach leverages the assumption that true causal dependencies are invariant across changing operational or environmental conditions, even as observed correlations fluctuate.

Diffusion- and Variational-Based Graph Construction

In the forecasting paradigm (Liang et al., 2023), the encoder outputs known-mean and variance latent states per node. The decoder treats the evolution as Markovian in latent space according to linear SDEs:

$dZ^t = F(t) Z^t dt + L(t) d\beta(t)$

The transition probability is Gaussian:

$$p_\theta(Z^t|Z^{t-1}) = \mathcal{N}(Z^t; \Psi Z^{t-1}, \Sigma_\theta)$$

Causal graph adjacency entries $A^\text{causal}_{ij}$ are constructed as the probability $p_\theta(z_j^t, z_i^{t-1})$ under this joint transition. This jointly models time-lagged node interactions, thus encoding causality via the graph edge probabilities.

3. Spatio-Temporal Sequence Encoding and Decoding

The spatial-temporal sequences are modeled with graph-aware recurrent units.

GCLSTM Encoder–Decoder

Each GCLSTM unit (Zhang et al., 3 Feb 2026) at time $k$ computes gates with graph convolutions:

$$f^{(k)} = \sigma(\text{GC}_f([\ x^{(k)}; h^{(k-1)}], A^{(k)}) + b_f),\quad \dots$$

with $$\text{GC}(Z, A) = D^{-\tfrac12}(A+I) D^{-\tfrac12} Z W_{GC}$$

The encoder processes time $k = t-w+1, \ldots, t$ producing $(h^{(t)}, c^{(t)})$; the decoder runs in reverse, generating reconstructions $\{\hat{x}^{(k)}\}$. This forms a sequence-to-sequence autoencoder over the graph.

Dynamic GCN Forecasting with Attention

For forecasting (Liang et al., 2023), dynamic adjacency matrices are constructed and used in dynamic GCN layers, whose outputs are subject to temporal attention mechanisms, producing spatially and temporally resolved output $\widehat{Y}_{t+1:t+T'}$.

4. Training Objectives and Loss Functions

CGSTAE utilizes a combination of reconstruction and regularization objectives depending on the stage.

• Reconstruction Loss: $L_{MSE}$ is used in all phases, measuring time-windowed sequence reconstruction.
• Causal Invariance (Process Monitoring): $$L_\text{invariance} (A; \{A^{(t)}\}) = \sum_{i,j} |A_{ij} - A^{(t)}_{ij}|$$
• Prior and Sparsity Regularizers: Penalize deviation from known edges and encourage sparse/discrete graphs.
• Variational ELBO (Forecasting): Evidence Lower Bound with KL divergence and transition likelihood:

$$ELBO_t = \mathbb{E}_{q_\phi(Z^t)} [\log p(A_{t+1} | Z^t)] - KL[q_\phi(Z^t |\cdots) \| p(Z^t)]$$

The reparameterization trick is adopted for gradient-based optimization with stochastic latent variables.

5. Interpretability and Uncertainty Quantification

CGSTAE architectures are designed to promote interpretability and explicit quantification of uncertainty.

• Causal Interpretability: The learned graph $A$ in (Zhang et al., 3 Feb 2026) is directly comparable to expert, literature, or process-based diagrams. Empirical studies demonstrate close alignment with domain ground truth, with most spurious or missing connections absent.
• Uncertainty Estimates: In the forecasting context, variational parameters ($\sigma^t$, $\Sigma_\theta$) quantify model uncertainty in node embeddings and edge existence (Liang et al., 2023).
• Diagnostic Visualization: 2D-Gaussian plots and violin-plots of learned joint embeddings make lagged or direct causation distinguishable via observed correlations.

A plausible implication is that such models allow for transparent, mechanism-aligned diagnosis, facilitating root cause tracing and process analysis.

6. Empirical Evaluation and Comparative Performance

CGSTAE models have been empirically validated on multiple domains:

• Process Monitoring (Zhang et al., 3 Feb 2026): On the Tennessee Eastman Process (TEP) and an argon-distillation air separation process (ASP), CGSTAE attains state-of-the-art F₁-scores (TEP: 0.896, ASP: 0.820) and low false alarm rates. Causal graphs learned on TEP align with three-quarters of known causal edges and remove nearly all spurious ones. In ASP, CGSTAE identifies root causes via minimal causal subgraphs.
• Spatio-Temporal Forecasting (Liang et al., 2023): On PeMS08, Los-loop, T-Drive, and synthetic fMRI datasets, CGSTAE yields lowest RMSE/MAE and recovers more accurate time-varying causal graphs (F1 increases of 10–20% over baselines such as VGAE, TCDF).
• Noise Robustness: Performance degrades minimally under Poisson noise ($\lambda=1,3,6$) (Liang et al., 2023).

A summary of comparative empirical results is provided:

Dataset / Metric	CGSTAE F₁	Best Baseline F₁	Comment
TEP (Process Mon.)	0.896	0.822 (DGSTAE)	FDR 0.822, FAR 0.059
ASP (Process Mon.)	0.820	0.784 (GAE-II)	FDR 0.941, FAR 0.057
PeMS08, fMRI, ...	↑	-	Lowest RMSE/MAE, robust causal recovery

7. Interpretive Insights and Applications

CGSTAE's explicit causal structure enables:

• Transparent anomaly/fault flagging in industrial monitoring, using Hotelling’s $T^2$ and squared prediction error (SPE) statistics against learned thresholds.
• Minimal causal subgraph extraction for root cause localization in multivariate systems.
• Dynamic, uncertainty-aware spatio-temporal predictions sensitive to both endogenous dynamics and exogenous disruptions.

The integration of spatial self-attention, GCLSTM, variational and diffusion-based decoding, together with causal invariance principles, establishes CGSTAE as a reference framework for interpretable, robust, and practical graph-based modeling in scientific and industrial domains (Liang et al., 2023, Zhang et al., 3 Feb 2026).