ST-Conv Block in Spatio-Temporal GCN

• ST-Conv Block is a structured module in spatio-temporal graph convolutional networks that fuses temporal and spatial convolutions to process multivariate time series.
• It sequentially applies a temporal convolution with GLU, a Chebyshev polynomial-based graph convolution, and a second temporal layer, adjusting tensor dimensions along the way.
• Its modular design facilitates efficient implementation and supports experimentation with hybrid temporal architectures for improved handling of complex temporal-graph data dependencies.

An ST-Conv Block is a structured building unit in Spatio-Temporal Graph Convolutional Networks (ST-GCN), integrating both temporal and spatial convolutions to model multivariate time series on graph-domain data. It fuses a 1D temporal convolution (optionally with gating), a graph convolution (realized via Laplacian and Chebyshev polynomial expansions), and a second temporal convolution, collectively operating on input of shape $X \in \mathbb{R}^{N \times C_{in} \times T}$, where $N$ is the number of nodes, $C_{in}$ is the input channel count, and $T$ is the temporal sequence length. This structural design supports implementation in any deep learning framework and serves as the core architectural component for ST-GCNs as formalized in Yu et al. and empirically analyzed in Turner (“Spatio-Temporal Graph Convolutional Networks: Optimised Temporal Architecture”) (Turner, 14 Jan 2025).

1. Architecture and Schematic

The ST-Conv Block comprises three serial sub-blocks, processing the tensor $X$ via temporal and spatial transformations. Denoting $C_h$ as the number of hidden channels, $K_t$ as temporal kernel size, and $K_s$ as the order of Chebyshev truncation:

$X$8

Typically, "valid" convolution yields $T_1 = T - (K_t-1)$ and $T_2 = T_1 - (K_t-1)$; zero-padding may be employed to maintain $N$0.

2. Mathematical Foundations

2.1 Spatial Graph Convolution

For adjacency $N$1, the normalized Laplacian is constructed as: $N$2 with $N$3 and degree matrix $N$4. Rescaling for Chebyshev polynomial approximation: $N$5 Using the Chebyshev polynomials $N$6 (with $N$7), for each time slice $N$8: $N$9 For $C_{in}$0, this reduces to: $C_{in}$1

2.2 Temporal 1D Convolution

For each node $C_{in}$2, channel $C_{in}$3, and time $C_{in}$4: $C_{in}$5 When using Gated Linear Unit (GLU) activation: $C_{in}$6 with elementwise gating via sigmoid $C_{in}$7. The second temporal layer can use either ReLU: $C_{in}$8 or GLU as above.

2.3 Block Composition

The full ST-Conv Block operation can be concisely written: $C_{in}$9 where each sub-operation may alter the time or channel dimensions in sequence.

3. Tensor Shapes and Dimensionality

The core tensor dimensions as they propagate through the block:

Stage	Shape	Comments
Input	$T$0	Original signal
H1 (Temp1)	$T$1	$T$2
H2 (GCN)	$T$3	Time unchanged by GCN
H_out (Temp2)	$T$4	$T$5

In Yu et al. and Turner, typical settings are $T$6, $T$7, $T$8, $T$9, yielding $X$0 and $X$1. To preserve input length, padding of $X$2 is applied. After two convolutions with kernel size $X$3, each output step depends on a local temporal window of size $X$4 centered at each time point; spatially, $X$5 considers one-hop graph neighbors.

4. Implementation and Forward Pass

The block is efficiently implementable in all modern frameworks. The core logic, written in pseudocode, is:

$X$9

A complete ST-GCN stacks two such blocks sequentially, then applies a final linear layer pointwise across nodes and time.

5. Rationale for CNN-Based Temporal Blocks and Alternatives

Yu et al. argue that CNN-based temporal blocks enable parallel training over the full time-axis, bypassing recurrent step-by-step dependencies of LSTM, employ simpler gating (via GLU), allow parameter sharing across all vertices prior to spatial convolution, and support efficient feature extraction using a learned $X$6-dimensional filter space. These design features yield significant practical advantages for training speed and model interpretability.

However, Turner et al. hypothesize that use of convolutional temporal blocks with fixed $X$7 can result in over-parameterization and increased risk of overfitting, especially when stacked with GCNs. The empirical findings in (Turner, 14 Jan 2025) indicate that replacing or augmenting temporal CNNs with LSTM blocks — or combining both within the block — often results in improved generalization, particularly for datasets exhibiting high noise or complex temporal dependencies. The hybrid CNN-GCN-LSTM system frequently outperforms pure CNN-based versions, suggesting task-dependent trade-offs.

6. Significance and Extensions

The ST-Conv Block provides a modular design for extracting robust spatiotemporal features from graph-based time series. Its explicit formalization (temporal convolutions, Chebyshev spectral GCN, and activation functions) enables reproducibility and simple extension to more elaborate deep learning architectures. The modular structure facilitates systematic benchmarking of CNN, LSTM, and hybrid temporal mechanisms within graph-based models. A plausible implication is that such modularity will enable more domain-adapted architectures, particularly as tasks and graph structures become more heterogeneous. The empirical investigations in (Turner, 14 Jan 2025) support ongoing hybridization of temporal modeling approaches to optimize both statistical and computational efficiency.