Spatio-Temporal Graph Neural Networks for Predictive Learning in Urban Computing: A Survey

In plain words

Cities produce large amounts of data that change over time and across places: traffic counts on roads, bike pick-ups at stations, air quality at monitors, crime and disease reports by district. Predicting how these numbers will move helps with planning, such as sending ambulances or police where they will be needed.

Many of these sensors and places are linked. A busy road affects the roads next to it, and a neighbouring district often looks similar. A graph is a natural way to store those links: each sensor is a node, and the edges say which sensors relate. A spatio-temporal graph neural network (STGNN) uses such a graph to mix information between neighbours, and also tracks how each node changes over time.

This survey does not propose a new model. It maps the field: how to build the graph from roads, distances, similarity or flows; which neural parts are used for the spatial side and the time side; how the two are combined; the newer variants that learn the graph or split the signal into parts; and the open problems such as interpretability, uncertainty, distribution shift and scale. Most of the work it covers is about traffic.

TL;DR

A survey of STGNNs for predictive learning in urban computing. It sorts the field into application domains, graph construction methods, basic building blocks (graph convolution, GRU, gated and causal TCN, self-attention) with factorized and coupled fusion, and recent variants and advanced frameworks §1 fig. 3. The paper runs no experiments and gives no code of its own. Its only numbers are literature counts: transportation is over 60% of the surveyed applications §4 fig. 4. Use it as a map and as a list of papers to read, not as evidence that one model is better than another.

Problem

Urban spatio-temporal data are correlated in space and in time, and heterogeneous across places and time ranges §1. Classic models such as SVR, random forests and gradient-boosted trees struggle to capture this complex structure, and CNN/RNN hybrids cannot directly take non-Euclidean inputs such as road networks or knowledge graphs §1. The survey asks how STGNNs, which combine graph learning with temporal learning, are built and used for prediction in urban settings §1. It also sets out what is missing: interpretability, calibration of uncertainty, physical constraints, pre-training, fine-grained architecture search, robustness to distribution shift, training strategies and scalability §8.

Method

This is a survey, so the "method" is its taxonomy. Sections below follow the compiled paper's numbering: §2 graph construction, §3 taxonomy, §4 application domains, §5 basic neural architectures, §6 STGNN variants, §7 advanced learning frameworks, §8 challenges, §9 conclusion §1 fig. 3.

Framework §5 fig. 5

The paper describes an STGNN as three modules: a data processing module that builds the spatio-temporal graph from raw data, a spatio-temporal graph learning module (STGLM) that captures the spatio-temporal dependencies, and a task-aware prediction module (TPM) that maps the hidden representation to the target. The STGLM is called the most important part §5 fig. 5.

# Basic STGNN for predictive learning  [§5] [fig. 5]
# Input:  raw sensor observations over a window
# Output: predictions for the next steps, one row per node

def stgnn_forecast(raw_observations, graph_source):
    # Data processing module: build the spatio-temporal graph  [§2] [§5]
    graph_sequence = build_graph(raw_observations, graph_source)
    # STGLM: spatial and temporal learning, joined by a fusion design  [§5.5]
    hidden = fuse_spatial_temporal(graph_sequence)
    # Task-aware prediction module: map hidden states to the targets  [§5]
    return predict_head(hidden)

The paper's general prediction form is that a window of past states plus the graph is mapped to a window of future states eq. 5:

\left[\mathcal X_{(t-T'+1)}, \cdots, \mathcal X_{(t)} ; \mathcal{G}\right] \xrightarrow{f(\cdot)} \left[\mathcal X_{(t+1)}, \cdots, \mathcal X_{(t+T)}\right]

eq. 5 §4.1

Spatio-temporal graph construction §2

A spatio-temporal graph is G_t = (V, E_t, A_t) with N vertices, where the edge set and adjacency may change over time. Graphs may be directed or undirected, weighted or unweighted, static or dynamic §2 fig. 2. The paper groups predefined constructions into four classes §2:

• Topology-based: unweighted 0/1 adjacency from a given structure such as a road network eq. 1 §2.
• Distance-based: a kernel of the distance between nodes, cut off at a threshold eq. 2 §2.
• Similarity-based: correlation of time series (Pearson correlation coefficient; dynamic time warping is named as the other common option) or of attributes such as points of interest eq. 3 §2.
• Interaction-based: each flow from node i to j normalised by the total flow out of i to its interacting nodes eq. 4 §2.

Adaptive graphs, learned instead of predefined, are described in §6.1.2.

a_{ij}^{t}=\begin{cases} 1, & \text{if } v_i \text{ connects to } v_j \\ 0, & \text{otherwise} \end{cases}

eq. 1 §2

a_{ij}^{t}=\begin{cases} \dfrac{\exp\left(-\lVert d_{ij}^{t}\rVert_2\right)}{\sigma}, & \text{if } d_{ij}^{t}<\epsilon \\ 0, & \text{otherwise} \end{cases}

eq. 2 §2 (as printed; see Pitfalls)

a_{ij}^{t}=\begin{cases} \dfrac{\sum_{n=1}^{n_s}\left(x_i^{0:t}-\bar{x}_i\right)\left(x_j^{0:t}-\bar{x}_j\right)} {\sqrt{\sum_{n=1}^{n_s}\left(x_i^{0:t}-\bar{x}_i\right)^2}\, \sqrt{\sum_{n=1}^{n_s}\left(x_j^{0:t}-\bar{x}_j\right)^2}}, & \text{(condition not stated)} \\ 0, & \text{otherwise} \end{cases}

eq. 3 §2 (sum index written as n here; the paper reuses i inside the sum. The "if" condition is not stated; see Pitfalls)

a_{ij}^{t}=\begin{cases} \dfrac{F_{ij}^{t}}{\sum_{m\in N(i)} F_{im}^{t}}, & \text{if } F_{ij}^{t}>0 \\ 0, & \text{otherwise} \end{cases}

eq. 4 §2

Basic building blocks §5

Graph convolution §5.1. The spectral form applies a filter in the Laplacian eigenbasis eq. 6 §5.1.1. The paper names ChebNet, a truncated Chebyshev expansion of the filter, as a popular choice §5.1.1. The spatial form, from Kipf and Welling, is eq. 7 §5.1.2:

\boldsymbol{g}_w \star \boldsymbol{x} = \boldsymbol{U}\boldsymbol{g}_w\boldsymbol{U}^T\boldsymbol{x}

eq. 6 §5.1.1

\boldsymbol{g}_w \star \boldsymbol{x} = w\left(\boldsymbol{I}_N+\boldsymbol{D}^{-\frac{1}{2}}\boldsymbol{A}\boldsymbol{D}^{-\frac{1}{2}}\right)\boldsymbol{x}

eq. 7 §5.1.2

The graph attention network §5.1.3 scores each neighbour u of node v and aggregates eq. 8:

\begin{aligned} \boldsymbol{h}_v^{t+1} &= \rho\!\left(\sum_{u \in \mathcal{N}_v} \alpha_{vu}\,\boldsymbol{W}\boldsymbol{h}_u^{t}\right) \\ \alpha_{vu} &= \frac{\exp\!\left(\operatorname{LeakyReLU}\!\left(\boldsymbol{a}^T\left[\boldsymbol{W}\boldsymbol{h}_v \,\|\, \boldsymbol{W}\boldsymbol{h}_u\right]\right)\right)} {\sum_{k \in \mathcal{N}_u} \exp\!\left(\operatorname{LeakyReLU}\!\left(\boldsymbol{a}^T\left[\boldsymbol{W}\boldsymbol{h}_v \,\|\, \boldsymbol{W}\boldsymbol{h}_k\right]\right)\right)} \end{aligned}

eq. 8 §5.1.3 (as printed; see Pitfalls)

Recurrent units §5.2. The text names vanilla RNN, LSTM and GRU, and says GRU is the most widely used variant. The GRU is given with an update gate u_t, a reset gate r_t, a candidate state and a mixing step eq. 9 §5.2:

\begin{aligned} \boldsymbol{u}_t &= \sigma(\boldsymbol{W}_u x_t + \boldsymbol{U}_u \boldsymbol{C}_{t-1} + \boldsymbol{b}_u) \\ \boldsymbol{r}_t &= \sigma(\boldsymbol{W}_r x_t + \boldsymbol{U}_r \boldsymbol{C}_{t-1} + \boldsymbol{b}_r) \\ \tilde{\boldsymbol{C}}_t &= \tanh(\boldsymbol{W}_C x_t + \boldsymbol{U}_C(\boldsymbol{r}_t \odot \boldsymbol{C}_{t-1}) + \boldsymbol{b}_C) \\ \boldsymbol{C}_t &= \boldsymbol{u}_t \odot \boldsymbol{C}_{t-1} + (1-\boldsymbol{u}_t) \odot \tilde{\boldsymbol{C}}_t \end{aligned}

eq. 9 §5.2

Temporal convolution §5.3. A gated TCN multiplies a tanh branch by a sigmoid-type gate eq. 10 §5.3.1. A causal, dilated 1D convolution reads only past inputs, with dilation d eq. 11 §5.3.2.

\mathcal{F}(x) = \tanh(\boldsymbol{\Theta}_1 \star x) \odot \sigma(\boldsymbol{\Theta}_2 \star x)

eq. 10 §5.3.1

F(s) = (\boldsymbol{x} *_d f)(s) = \sum_{i} f(i)\, x_{s-d\cdot i}

eq. 11 §5.3.2 (lower summation limit not printed; see Pitfalls)

Temporal self-attention §5.4. Scaled dot-product attention with Q, K, V obtained by linear maps of the input, plus positional information eq. 12 §5.4:

\operatorname{Attention}(\boldsymbol{Q}, \boldsymbol{K}, \boldsymbol{V}) = \operatorname{Softmax}\!\left(\frac{\boldsymbol{Q}\boldsymbol{K}^{T}}{\sqrt{d_k}}\right)\boldsymbol{V}

eq. 12 §5.4

Spatio-temporal fusion §5.5. The paper distinguishes two designs:

• Factorized §5.5.1: spatial and temporal layers are stacked in series or in parallel. STGCN is given as a sandwich of two temporal convolutions around one graph convolution eq. 13 §5.5.1 fig. 6. T-GCN runs a GCN, then a GRU, at each time step eq. 14 §5.5.1 fig. 7.
• Coupled §5.5.2: the graph convolution replaces the linear maps inside a recurrent cell. DCRNN is the example eq. 15 §5.5.2 fig. 8.

v^{l+1}=\boldsymbol{\Gamma}_1^{l} *_{\mathcal{T}} \operatorname{ReLU}\!\left(\boldsymbol{\Theta}^{l} *_{\mathcal{G}} \left(\boldsymbol{\Gamma}_0^{l} *_{\mathcal{T}} v^{l}\right)\right)

eq. 13 §5.5.1

\begin{aligned} \boldsymbol{f}(\boldsymbol{X}, \boldsymbol{A}) &= \sigma\!\left(\boldsymbol{A}\boldsymbol{X}\boldsymbol{W}_0\right) \\ \boldsymbol{u}_t &= \sigma\!\left(\boldsymbol{W}_u\left[\boldsymbol{f}(\boldsymbol{A}, \mathcal{X}_t), \boldsymbol{H}_{t-1}\right] + \boldsymbol{b}_u\right) \\ \boldsymbol{r}_t &= \sigma\!\left(\boldsymbol{W}_r\left[\boldsymbol{f}(\boldsymbol{A}, \mathcal{X}_t), \boldsymbol{H}_{t-1}\right] + \boldsymbol{b}_r\right) \\ \boldsymbol{c}_t &= \tanh\!\left(\boldsymbol{W}_c\left[\boldsymbol{f}(\boldsymbol{A}, \mathcal{X}_t), \left(\boldsymbol{r}_t * \boldsymbol{H}_{t-1}\right)\right] + \boldsymbol{b}_c\right) \\ \boldsymbol{H}_t &= \boldsymbol{u}_t * \boldsymbol{H}_{t-1} + (1-\boldsymbol{u}_t) * \boldsymbol{c}_t \end{aligned}

eq. 14 §5.5.1 (as printed; see Pitfalls)

\begin{aligned} \boldsymbol{r}_t &= \sigma\!\left(\boldsymbol{\Theta}_r \star \mathcal{G}[\mathcal{X}_t, \boldsymbol{H}_{t-1}] + \boldsymbol{b}_r\right) \\ \boldsymbol{u}_t &= \sigma\!\left(\boldsymbol{\Theta}_u \star \mathcal{G}[\mathcal{X}_t, \boldsymbol{H}_{t-1}] + \boldsymbol{b}_u\right) \\ \boldsymbol{C}_t &= \tanh\!\left(\boldsymbol{\Theta}_C \star \mathcal{G}[\mathcal{X}_t, (\boldsymbol{r}_t \odot \boldsymbol{H}_{t-1})] + \boldsymbol{b}_c\right) \\ \boldsymbol{H}_t &= \boldsymbol{u}_t \odot \boldsymbol{H}_{t-1} + (1-\boldsymbol{u}_t) \odot \boldsymbol{C}_t \end{aligned}

eq. 15 §5.5.2

Variants §6

Spatial learning §6.1:

• Multi-graph convolution §6.1.1 fig. 9: several graphs (neighbourhood, function similarity, connectivity) each get their own graph convolution; STMGCN combines a contextual gated RNN with ChebNet per graph. No equation is given.
• Adaptive graph learning §6.1.2: the graph is learned. Graph WaveNet builds the adjacency from two learnable node-embedding matrices eq. 16. MTGNN builds it from tanh-transformed embeddings, with a saturation hyperparameter α eq. 17, eq. 18. Feature-based variants (e.g. DGCRN, GTS) build the graph from inputs or hidden states; no equations for them are given.
• Multi-scale spatial learning §6.1.3: nodes are grouped into communities (by hand, by clustering, or learned), with inter- and cross-community correlations. ST-SHN and ST-HSL learn hyperedges; GAGNN learns a grouping; THINK and DMGCRN use hyperbolic graph networks. No equations.
• Heterogeneous spatial learning §6.1.4: different edge types (static vs dynamic, transport mode) get separate aggregation. No equations.

\tilde{\boldsymbol{A}}_{adp} = \operatorname{SoftMax}\!\left(\operatorname{ReLU}\!\left(\boldsymbol{E}_1 \boldsymbol{E}_2^{T}\right)\right)

eq. 16 §6.1.2 — \boldsymbol{E}_1, \boldsymbol{E}_2 \in \mathbb{R}^{N\times C} randomly initialised

\begin{aligned} \boldsymbol{M}_1 &= \tanh\!\left(\alpha \boldsymbol{E}_1 \boldsymbol{\Theta}_1\right), \qquad \boldsymbol{M}_2 = \tanh\!\left(\alpha \boldsymbol{E}_2 \boldsymbol{\Theta}_2\right) \\ \tilde{\boldsymbol{A}}_{adp} &= \operatorname{ReLU}\!\left(\tanh\!\left(\alpha\left(\boldsymbol{M}_1\boldsymbol{M}_2^{T} - \boldsymbol{M}_2\boldsymbol{M}_1^{T}\right)\right)\right) \end{aligned}

eq. 17 (M₁, M₂) and eq. 18 (Ã_adp) §6.1.2 — MTGNN

Temporal learning §6.2:

• Multi-scale temporal §6.2.1: TCNs with several receptive fields (MTGNN uses several kernel sizes); or combined TCN and Transformer for short- and long-range patterns (DMVST-VGNN).
• Multi-granularity §6.2.2 fig. 10: three branches for the latest hour, the same hour on the previous day, and the same hour on the previous week (ASTGCN), fused by a learnable weight matrix.
• Decomposition §6.2.3: the series is split into components, e.g. with N-BEATS-style residual subtraction (FC-GAGA, StemGNN), a temporal residual decomposition (D2STGNN), or a discrete wavelet transform (STWave). No equations.

Spatio-temporal fusion §6.3:

• Joint modelling §6.3.1: one graph over space and time. STSGCN links each node to itself at adjacent time steps and runs a GCN on the result; STFGNN adds similarity graphs; S2TAT uses attention. No equations.
• Automated fusion §6.3.2: neural architecture search. AutoSTG applies DARTS to stacked cells and uses meta-knowledge to generate weights; AutoSTS and Auto-DSTSGN search GCN/TCN choices and adjacency matrices. No equations.

Advanced frameworks §7

• Adversarial learning §7.1: a discriminator added to a standard loss (GAN-style), with the STGNN as generator. No equations.
• Meta learning §7.2 fig. 11: ST-MetaNet has node- and edge-level meta-knowledge learners that generate weights for its GRU and GAT parts. MegaCRN adds an attention memory over seen patterns. No equations.
• Self-supervised / contrastive §7.3 fig. 12: STGCL builds positive and negative samples by edge masking, input masking and temporal shifting, encodes both, and adds a contrastive loss (attributed to GraphCL) to a predictive MAE loss. Negatives are filtered by graph neighbourhood and periodicity. The loss formula is not in the compiled paper.
• Continuous modelling §7.4: a neural ODE over graph convolution (STGODE), jointly over space and time (MTGODE), or a neural CDE (STG-NCDE). MixRNN+ combines a neural ODE with an RNN. No equations.
• Physics-informed §7.5 fig. 13: STAN has GAT and GRU parts with SIR-consistent outputs (transmission rate, recovery rate, change in infections and recoveries) under a constraint. The constraint is not written in the compiled paper.
• Transfer learning §7.6: clustering-based (TL-DCRNN), domain-adaptation-based (DASTNet) and meta-learning-based transfer across cities. No equations.

Hyperparameters

This survey trains nothing. The table lists the parameters that appear in the equations and whether the paper gives a value.

Results

The paper reports no experiments. Its numbers are literature counts from its own search, and they are not benchmark results.

The paper does not list the keywords used for the Google Scholar counts, so the counts cannot be reproduced exactly.

Datasets tab. 2

Links are copied from the paper's table. Reference lists per dataset are not reproduced here; see tab. 2 in the paper.

Notation tab. 1

Paper vs code

No official code. The paper is a survey with no model of its own. Its TeX source contains a commented-out table of source-code links for about 40 models. That table is not in the compiled paper and is not used here. A GitHub search for the survey title and for "STGNN survey" with the authors' names found no repository.

Pitfalls

• No experiments, no code, no training details. Nothing here shows that one architecture beats another. The card's numbers are literature counts only.
• Inconsistent feature dimension. §2 writes node features as N×F; the prediction form eq. 5 writes N×d. Pick one and state which the code uses.
• Operator clash. §5.1.1 says ⊙ is the convolution operator. Table 1 defines ⋆ as convolution and ⊙ as element-wise product, and the equations follow Table 1. Read the text's ⊙ as a typo.
• Gaussian kernel form eq. 2. The text calls it a Gaussian radial basis function, but the printed formula is exp(−‖d‖₂)/σ, without a square or σ² on the distance. σ and ε are not given. Do not assume the textbook Gaussian without checking the cited source.
• PCC graph eq. 3. The "if" branch has no condition, so the threshold is missing. The sum reuses index i for both nodes and time samples. Dynamic time warping is written as "wrapping" in the text.
• Spatial GCN eq. 7. The operator is I + D⁻½AD⁻½. The paper does not say whether A includes self-loops, whether D includes them, or whether w is a scalar or a matrix.
• GAT eq. 8. The numerator sums over u ∈ N_v, but the denominator sums over k ∈ N_u. This is probably a typo for N_v; the card cannot confirm it. ρ and the LeakyReLU slope are not given.
• Causal convolution eq. 11. The lower summation limit is not printed (it reads Σ_i^{k−1}); the range has to be set from the kernel size. The prose says causal convolution "removes connections between future and past time steps", which is ambiguous. The formula reads only past inputs x_{s−d·i}, so follow the formula.
• Attention scale eq. 12. The formula divides by √d_k, but the text says the scale "is equal to the dimension of the model". The paper does not say whether d_k is per-head or the model width. Positional encodings are named but not written in the compiled text.
• DCRNN eq. 15 vs text. The text says the linear units inside an LSTM are replaced, but the equations are GRU-form (update and reset gates). Implement the equations.
• T-GCN eq. 14. The definition is written f(X, A) but used as f(A, 𝒳_t). σ is not defined in the text.
• MTGNN eq. 17. The paper writes θ₁, θ₂ in the text and Θ₁, Θ₂ in the equation. α, C and the initialisation are not stated.
• STGCL §7.3. The contrastive loss and sampling details are not in the compiled text; only the augmentation names and a GraphCL citation are. Do not implement the loss from this card.
• STAN §7.5. The physical constraint is not written in the compiled text.
• Third-party figures. A footnote says that figures not drawn by the authors are used "with the permission of their authors". Captions credit STGCN, T-GCN, DCRNN, STMGCN, ASTGCN, ST-MetaNet, STGCL, STAN and a temporal-shift figure from another paper. This CC BY 4.0 licence covers only the authors' own work, so none of these figures is used here.
• Counts without method. "Over 60%" and the publication counts do not say which papers were counted or how the keywords were chosen.
• Dataset table as printed. Some links have stray spaces or typos (e.g. "sytem-data", "dcjingsai.com./"). They are copied as printed and were not checked.

Lineage

• Builds on: none. It surveys the STGNN literature, including the GNN, GRU, TCN and Transformer papers it cites.
• Related surveys named in the paper, without a card here: Wang et al. (2020) on deep spatio-temporal data mining; Ye et al. (2020) on urban traffic; Bui et al. (2022) and Jiang et al. (2022) on traffic forecasting; Sahili et al. (2023) and Li et al. (2023) on STGNNs. Their content is not used on this card.
• Superseded by: none known from content in this wiki.