TK-GCN: Diverse Graph Convolution Approaches

• TK-GCN is an ambiguous acronym representing multiple graph convolution variants, including TrafficKAN-GCN, Taylor-kernel GCN, and Transformer with Koopman-enhanced GCN.
• TrafficKAN-GCN integrates Kolmogorov–Arnold Networks with GCN to optimize urban traffic dynamics, reducing peak delays and supporting adaptive route control.
• Taylor-based and Transformer-enhanced formulations use advanced spectral filtering and latent linearization, improving forecasting on irregular spatiotemporal domains.

TK-GCN is an acronym used for several non-identical graph neural network architectures rather than a single canonical model family. In the provided literature, it denotes or is associated with at least three distinct directions: TrafficKAN-GCN, a hybrid of Kolmogorov–Arnold Networks and Graph Convolutional Networks for urban traffic flow optimization; Taylor-based GCN, also described as TK-GCN in the sense of a Taylor-kernel GCN derived from Taylor approximation of spectral wavelet-kernels; and Transformer with Koopman-Enhanced GCN, a two-stage framework for long-horizon spatiotemporal forecasting on irregular domains. A related tensor/Kronecker line, represented by TGCNN, is presented as relevant to TK-GCN in a Kronecker/tensor sense (Zhang et al., 5 Mar 2025, Zhang et al., 2020, Wang et al., 5 Jul 2025, Zhang et al., 2018).

1. Terminological scope

The acronym’s ambiguity is central to understanding the literature. The same short form attaches to different augmentations of graph convolution, and the technical content changes substantially from one paper to another.

Usage	Paper	Defining mechanism
TrafficKAN-GCN	"TrafficKAN-GCN: Graph Convolutional-based Kolmogorov-Arnold Network for Traffic Flow Optimization" (Zhang et al., 5 Mar 2025)	KAN blocks replace fixed activations inside a GCN pipeline
Taylor-kernel / Taylor-based GCN	"From Spectrum Wavelet to Vertex Propagation: Graph Convolutional Networks Based on Taylor Approximation" (Zhang et al., 2020)	Taylor approximation of spectral wavelet-kernels in vertex-domain propagation
Transformer with Koopman-Enhanced GCN	"Transformer with Koopman-Enhanced Graph Convolutional Network for Spatiotemporal Dynamics Forecasting" (Wang et al., 5 Jul 2025)	Koopman-enhanced latent encoding followed by masked Transformer forecasting
Tensor/Kronecker-related lineage	"Tensor graph convolutional neural network" (Zhang et al., 2018)	Parameterized Kronecker-sum cross-graph convolution on factorizable graphs

A common misconception is to treat TK-GCN as a single established architecture. The literature summarized here does not support that interpretation. Instead, the acronym functions as a compact label for different ways of extending graph convolution: adaptive nonlinear approximation, Taylor polynomial propagation, latent linearization with attention, or tensorized cross-graph composition. This suggests that the full paper title or arXiv identifier is necessary for disambiguation in technical discussion.

2. TrafficKAN-GCN: KAN–GCN coupling for urban traffic optimization

TrafficKAN-GCN, abbreviated in the paper as TK-GCN, is a hybrid graph-learning framework that integrates Kolmogorov–Arnold Networks (KAN) with Graph Convolutional Networks (GCN) to model and optimize urban traffic flows under complex, nonlinear, and time-varying spatiotemporal dynamics (Zhang et al., 5 Mar 2025). Its motivation is explicitly framed against static shortest-path formulations: Dijkstra’s algorithm assumes fixed arc weights and does not learn from data; Floyd–Warshall’s $O(n^3)$ runtime is impractical for real-time global optimization over large urban graphs; and Genetic Algorithms are flexible but slow to converge and computationally heavy.

The framework constructs an undirected weighted graph $G=(V,E,W)$, where nodes are intersections and transit hubs, and edges connect adjacent nodes. Edge weights combine length, speed, congestion, and historical travel time:

$$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$

with trainable $\alpha,\beta,\gamma,\delta$. Adjacency normalization is defined as

$$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$

The GCN update used in the paper is

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$

and the distinctive modification is that fixed MLP activations are replaced by KAN blocks. The KAN component follows the Kolmogorov–Arnold style parametrization

$$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$

with learnable univariate functions $g_i$ and $\phi_j$, implemented through spline or basis expansions. The experiments report that grid size $=1$ and spline order $G=(V,E,W)$0 yield the best accuracy on both the Baltimore traffic and CORA datasets.

The data pipeline aggregates road network topology, AADT traffic volume counts, derived congestion levels from real-time feeds, public transit routes and schedules, and GIS layers including lane count and road type. Preprocessing rules are explicit: missing values are imputed with KNN if under 5% and discarded if over 30%; outliers are filtered by Z-score with $G=(V,E,W)$1 and then smoothed by a rolling mean; and all features are min–max normalized to $G=(V,E,W)$2. Temporal handling is feature-based rather than recurrent: moving averages

$G=(V,E,W)$3

and cyclical hour-of-day encodings

$G=(V,E,W)$4

are concatenated to node inputs. The paper states explicitly that there is no explicit recurrent module.

Training uses a composite objective,

$G=(V,E,W)$5

with

$G=(V,E,W)$6

Optimization uses Adam with a decaying learning rate, batch-wise graph updates, typically 300 epochs, batch size 64, and 5-fold cross-validation.

On Baltimore traffic, the reported model comparison is as follows:

Model	MAE / RMSE	Training
MLP-GCN	3.52 / 4.89	124.5 s, 180 epochs
TK-GCN (KAN-GCN)	3.61 / 5.02	217.8 s, 250 epochs
Standard GCN	3.75 / 5.20	152.3 s, 200 epochs
Transformer	3.47 / 4.80	390.1 s, 220 epochs

The same section also reports Dijkstra: MAE 4.00, RMSE 5.30 and GA: MAE 4.10, RMSE 5.40. The paper’s interpretation is specific: MLP-GCN delivers the best accuracy and efficiency; TK-GCN is competitive but slower due to KAN’s adaptive functions and graph computations; Transformer is accurate for long-horizon forecasting but incurs the largest training cost.

The Baltimore case study centers on the disruption caused by the Francis Scott Key Bridge collapse. Reported outcomes include peak-hour delays reduced by 22% after optimization and low-traffic bus stops decreased by 38% while high-utilization hubs increased proportionally. The paper states that traffic shifted to I-95, I-895, and MD-295, and that TK-GCN-based predictions support reweighting routes and control policies, including new bridges, roundabouts, and adaptive signals. Limitations are also explicit: computational overhead, overfitting risk when feature diversity is limited, absence of richer temporal modeling, and lack of live city-scale deployment.

3. Taylor-kernel GCN: Taylor approximation as graph propagation

In the Taylor-based line, the paper uses “Taylor-based GCN (TGCN)”, while the provided synthesis notes that TK-GCN (Taylor-kernel GCN) is an equivalent descriptive name emphasizing the kernel view (Zhang et al., 2020). The method revisits the derivation of graph convolution from spectral wavelet-kernels and replaces the usual first-order Chebyshev-style propagation with Taylor-expanded vertex-domain operators.

The central theoretical statement is that if a wavelet-kernel $G=(V,E,W)$7 is polynomial, or analytic on a domain containing the spectrum of a symmetric propagation matrix $G=(V,E,W)$8, then the spectral filter can be written as a polynomial in $G=(V,E,W)$9. The truncated Taylor form is

$$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$0

The general layer is expressed as

$$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$1

The paper then defines four variants. Type-1 is a first-order model with scalar center $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$2:

$$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$3

Type-2 replaces the scalar by a diagonal center $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$4:

$$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$5

Type-3 and Type-4 extend to higher-order propagation, with shared or per-order weights respectively.

The method’s stated advantages are flexibility in graph representation and learnable self-influence. The paper analyzes $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$6, $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$7, $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$8, and PageRank-style operators, and reports that normalized adjacency $$w_{ij} = \alpha L_{ij} + \beta S_{ij} + \gamma C_{ij} + \delta T_{ij},$$9 consistently outperformed raw $\alpha,\beta,\gamma,\delta$0 and was on par with or better than $\alpha,\beta,\gamma,\delta$1. It also reports that optimal $\alpha,\beta,\gamma,\delta$2 in Type-1 typically lies in $\alpha,\beta,\gamma,\delta$3.

On the random subset citation-network split, the reported results for first-order methods are: baseline GCN at 78.7±2.7% on Cora, 68.7±2.7% on Citeseer, and 78.8±3.1% on Pubmed; Type-2 TGCN at 81.3±2.7%, 70.3±2.6%, and 79.8±2.6% respectively. On the 60/20/20 split, Type-2 TGCN reports 87.23 on Cora, 78.31 on Citeseer, and 86.89 on Pubmed, compared with APPNP 86.88 / 77.74 / 88.41, GAT 86.37 / 74.32, and GCN 85.77 / 88.13 where listed. The paper therefore positions the method as competitive or better on Cora and Citeseer and slightly behind APPNP on Pubmed.

The same work extends beyond citation graphs. On ShapeNet point cloud segmentation, mean accuracies for TGCN Type-1 and Type-2 are reported in the range ~0.783–0.797, versus ~0.779 for GCN and ~0.803 for PointNet. On synthetic SBM tasks, Type-2 TGCN reaches 85.87 on the Pattern task and 61.2 on the Cluster task, exceeding GCN on the latter. Training uses Adam, learning rate 0.01, weight decay $\alpha,\beta,\gamma,\delta$4, and dropout 0.5. Reported limitations include sensitivity to $\alpha,\beta,\gamma,\delta$5, oversmoothing and overfitting at larger orders, and the need to keep the spectrum of $\alpha,\beta,\gamma,\delta$6 bounded for numerical stability.

4. Transformer with Koopman-Enhanced GCN: TK-GCN for latent-space forecasting

A third usage of TK-GCN appears in “Transformer with Koopman-Enhanced Graph Convolutional Network for Spatiotemporal Dynamics Forecasting”, where TK-GCN is defined as a two-stage framework combining geometry-aware spatial encoding, Koopman-based latent linearization, and Transformer temporal modeling (Wang et al., 5 Jul 2025). The target domain is an irregular geometric graph, and the paper’s application is spatiotemporal cardiac dynamics forecasting on a 3D ventricular mesh with 1,094 nodes and 2,184 elements derived from the PhysioNet Computation in Cardiology Challenge 2007.

The graph is $\alpha,\beta,\gamma,\delta$7, with node states $\alpha,\beta,\gamma,\delta$8, adjacency $\alpha,\beta,\gamma,\delta$9, and edge attributes $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$0 representing normalized coordinate differences. Stage 1 is a Koopman-enhanced GCN (K-GCN). It uses SplineCNN over pseudo-coordinates, with continuous B-spline kernels

$$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$1

and neighborhood aggregation

$$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$2

Residual connections are used in the form

$$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$3

The encoder $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$4 maps $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$5 to latent states $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$6, a dense matrix $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$7 advances the latent dynamics approximately linearly,

$$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$8

and the decoder $$\tilde{A} = D^{-1/2} A D^{-1/2}, \qquad D_{ii}=\sum_j A_{ij}.$$9 reconstructs back to the signal domain.

Stage 1 is trained with three losses:

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$0

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$1

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$2

with total objective

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$3

Stage 2 freezes $H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$4 and trains a masked Transformer decoder on sliding windows $H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$5 of latent states, with sinusoidal positional encodings, two decoder layers, 4 attention heads, and window length $H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$6. The forecasting loss is mean-squared error in latent space:

$H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$7

The experimental data are generated from the two-variable Aliev–Panfilov model under three stimulation protocols: Protocol I (Healthy), Protocol II (Localized perturbation), and Protocol III (Remote perturbation). Training uses the first 1,200 time steps and testing the next 300. The baseline set comprises LSTM, DMD, Pure Koopman, and VAR in the main experiments, and the ablation set includes Transformer + GCN, Transformer Only, Transformer + CNN-AE, and LSTM + K-GCN.

The ablation table reports the following TK-GCN MSE values across forecast intervals $H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$8, $H^{(l+1)} = \sigma(\tilde{A} H^{(l)} W^{(l)}),$9, and $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$0: Protocol I: 0.00012, 0.00040, 0.00062; Protocol II: 0.00037, 0.00051, 0.00044; Protocol III: 0.00017, 0.00015, 0.00028. The paper states that TK-GCN retains the lowest median MSE and smallest variance in box-plot summaries, and that its advantage widens on the more challenging fragmented dynamics of Protocol III. It also attributes part of the gain to Koopman-consistent latent evolution: removing Koopman dynamics in Transformer + GCN raises mid- and long-horizon error substantially, while replacing geometry-aware GCN with CNN-AE degrades performance on irregular meshes.

The method’s limitations are also specified: approximate linear latent evolution may fail under extreme chaos or unmodeled inputs; the geometry is fixed; only the transmembrane potential $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$1 is forecast directly; and autoregressive rollout can accumulate drift over very long horizons.

5. Tensor/Kronecker-related lineage: TGCNN and factorized graphs

Although TGCNN is not presented as the same named model as the preceding variants, the provided synthesis treats it as relevant to TK-GCN in a Kronecker/tensor-based GCN sense (Zhang et al., 2018). Its target setting is factorizable graphs, especially sequential dynamic graphs and cross-attribute graphs. The architecture is built around a graph preserving layer consisting of cross graph convolution and graph pooling.

Given signals $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$2, $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$3 with adjacencies $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$4 and $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$5, the conjunctive signal is

$$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$6

The standard Kronecker sum is

$$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$7

and the paper’s learnable generalization is the parameterized Kronecker sum

$$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$8

where $$f(x_1,\dots,x_n)=\sum_{i=1}^{m} g_i\!\left(\sum_{j=1}^{n}\phi_j(x_j)\right),$$9, $g_i$0, and $g_i$1. Spectral filtering is then performed by adjacency polynomials,

$g_i$2

followed by graph pooling:

$g_i$3

For sequences of subgraphs, the graph preserving layer is applied recursively. The current subgraph and the previously preserved subgraph are crossed, filtered, and pooled to produce a compact preserved graph, which acts as memory for subsequent steps. In the paper’s formulation,

$g_i$4

followed by pooled filtered outputs and a nonlinear projection to $g_i$5.

The paper emphasizes efficiency. Instead of explicitly convolving on the $g_i$6 conjunctive adjacency, Kronecker algebra maps the polynomial filtering into smaller matrix compositions. The synthesis states that, intuitively, the cost goes from $g_i$7 toward $g_i$8. This is the technical basis for applying the model to dynamic skeleton graphs and matrix completion.

Reported results are strong within those settings. On NTU RGB+D, TGCNN achieves 71.4% accuracy on Cross Subject and 82.9% on Cross View. On LSC, it reports Precision 86.6% / Recall 82.9% for RCSam and Precision 83.1% / Recall 76.5% for RCSub. On the synthetic Netflix matrix completion benchmark, TGCNN reports RMSE 0.0042, compared with 0.0049 for the classic Kronecker sum variant. The paper also reports that recursive factorized modeling outperforms isolated per-graph convolution and that the parameterized Kronecker sum improves over the non-parameterized version.

The limitations are equally clear. The method is most effective when the target graph can be factorized into meaningful subgraphs with well-defined intra-subgraph adjacencies. The learnable diagonal parameterization is less expressive than a fully dense cross-factor mapping, and the spectral filtering uses adjacency polynomials without explicit normalization.

6. Comparative structure, recurring themes, and disambiguation

Across the surveyed literature, the shared substrate is graph convolution, but the added mechanism changes the model’s semantics. TrafficKAN-GCN inserts KAN blocks to improve nonlinear approximation on traffic graphs; the Taylor-kernel line redefines propagation through Taylor polynomial filters and learnable self-influence; the Koopman-enhanced line places graph convolution in a two-stage latent forecasting pipeline with a Transformer; and TGCNN builds a conjunctive graph across factored subgraphs using a parameterized Kronecker sum (Zhang et al., 5 Mar 2025, Zhang et al., 2020, Wang et al., 5 Jul 2025, Zhang et al., 2018).

The temporal role of the graph component also differs. TrafficKAN-GCN encodes time through moving averages and cyclical features and explicitly states that it has no explicit recurrent module. The Taylor-based formulation is not primarily temporal; it is evaluated mainly on citation networks, point cloud segmentation, and synthetic node-classification tasks. The Koopman-enhanced TK-GCN is explicitly spatiotemporal, with staged training and masked autoregressive decoding. TGCNN handles temporal structure only when the input can be represented as a sequence of factorized subgraphs.

The computational trade-offs are similarly non-uniform. TrafficKAN-GCN reports 217.8 s training time versus 124.5 s for MLP-GCN and 152.3 s for standard GCN. Taylor-based higher-order propagation increases time and can oversmooth or overfit as $g_i$9 grows. The Koopman-enhanced Transformer carries the usual attention cost $\phi_j$0, though the reported setting $\phi_j$1 keeps it manageable. TGCNN reduces the cost of explicit cross-graph construction through Kronecker algebra, but only under the factorization assumptions built into the method.

These differences make acronym-only references potentially misleading. A statement such as “TK-GCN outperforms GCN” is not well-formed without the paper title or arXiv identifier, because the underlying operator, objective, graph representation, temporal mechanism, and evaluation domain may all differ. This suggests that, in technical writing, TK-GCN should be treated as an ambiguous shorthand whose meaning is fixed only by context: traffic optimization, Taylor-kernel propagation, Koopman-Transformer forecasting, or tensor/Kronecker cross-graph convolution.