Class Node Graph Attention Network (CGAT)

• The paper introduces CGAT as a graph-based model that fuses a learnable global CLS node with attention convolutions for interpretable 3D shape classification.
• It leverages stacked GATv2 layers and directed CLS edges to ensure that only the CLS node influences the final prediction, producing clear attention maps.
• Empirical results in forensic dental age estimation show that combining curvature and centroid features yields competitive metrics (F1 ~0.76) and enhanced interpretability.

Searching arXiv for the specified CGAT-related papers to ground the article in current records. Class Node Graph Attention Network (CGAT) is a graph-based architecture for explainable 3D shape classification in which a mesh-derived graph is augmented with a global CLS node and processed with stacked GATv2-style attention convolutions; the final prediction is made only from the CLS node embedding, and the learned attention into that node is visualized to indicate which regions of the input graph contributed most to the decision (Buyukcakir et al., 9 Sep 2025). In the reported use case, CGAT is applied to Demirjian stage allocation for third molars derived from CBCT images, where transparency is treated as a methodological requirement because the task is a high-stakes forensic application (Buyukcakir et al., 9 Sep 2025).

1. Definition and architectural rationale

CGAT was introduced to address a specific limitation of graph-based 3D recognition systems: standard deep learning models for 3D shapes can be accurate but are often black boxes, whereas forensic dental age estimation requires explanations that can be compared with expert reasoning (Buyukcakir et al., 9 Sep 2025). The architecture therefore combines graph attention convolutions with a virtual CLS node that acts as a global information aggregator and a single locus for explanation.

The defining design decision is that all decision-relevant information is funneled through the CLS node. The CLS embedding is initialized as a learnable parameter, connected to the mesh graph, updated through stacked attention-based message passing, and then passed through dropout and an MLP head to produce class logits. Softmax yields class probabilities. Because the classifier reads only the CLS node, attention into that node can be interpreted as a class-level explanation rather than merely a collection of local neighbor-to-neighbor importance scores.

The paper explicitly relates this idea to the role of the $[CLS]$ token in BERT or ViT, while keeping the implementation within a graph-attention setting rather than a sequence or image-patch setting (Buyukcakir et al., 9 Sep 2025). A plausible implication is that CGAT should be understood less as a generic synonym for graph attention and more as a specific pooling-and-explanation mechanism centered on a global node.

2. Graph representation of 3D meshes and feature design

In the reported application, each sample is a triangular mesh of a third molar obtained from CBCT scans collected at Belgian centers, covering FDI elements 18, 28, 38, and 48 (Buyukcakir et al., 9 Sep 2025). Each mesh is treated as an undirected graph,

$G(\mathcal{V}, \mathcal{E}),$

with vertices as nodes and mesh adjacencies as edges. During training, a batch of meshes is represented as a block-diagonal disconnected graph.

The preprocessing pipeline is designed to reduce computational cost and geometric variability while preserving shape information. Original meshes had on average about 17,205 nodes and 103,225 edges. They were simplified with quadric mesh decimation to about $751 \pm 1$ nodes and $4,497 \pm 4$ edges, and then scaled to fit inside a unit sphere so that the model would focus on shape rather than absolute size (Buyukcakir et al., 9 Sep 2025).

The node-feature study evaluates three alternatives: local mean curvature, distance to centroid, and both features combined. These features are described as pose-invariant and shape-informative. Mean curvature captures local concavity and convexity; the paper associates it with phenomena such as changes in root apices, furcations, and root formation. Distance to centroid captures the spatial elongation of the tooth, reflecting that early teeth tend to be more compact or rounded whereas later teeth elongate as roots develop. All features were min-max scaled into $[-1,1]$ (Buyukcakir et al., 9 Sep 2025).

This feature design is tightly coupled to the stated interpretability goal. The paper reports that the combined use of local mean curvature and distance to centroid not only yielded a slight performance increase but also more comprehensive attention visualizations (Buyukcakir et al., 9 Sep 2025). This suggests that the node-feature choice affects explanation quality as much as predictive quality.

3. Attention mechanism, CGAT blocks, and the role of the CLS node

CGAT is built on GATv2-style attention convolutions. The paper first recalls a standard GAT layer: ${h_i}^{(l+1)} = \concat_{k=1}^K \sigma\left(\sum_{j : j \in \mathcal{N}_i}\alpha^{(l,k)}_{ij} \boldsymbol{W}^{(l,k)} h_i^{(l)}\right),$ with attention coefficients

$$\alpha^{(l,k)}_{ij} = \mathrm{softmax}(e(h_i,h_j)),$$

$$e(h_i,h_j) = \text{LeakyReLU}\left(\boldsymbol{\alpha}_{(l,k)}^\top \left[\boldsymbol{W}^{(l,k)}h_i \,||\, \boldsymbol{W}^{(l,k)}h_j \right]\right).$$

It then motivates GATv2 by stating that original GAT attention is more static, whereas GATv2 changes the order of operations: $$e(h_i,h_j) = \boldsymbol{\alpha}_{(l,k)}^\top \text{LeakyReLU}\left( \boldsymbol{W}^{(l,k)}\cdot\left[h_i||h_j\right]\right),$$ which is presented as more dynamic and better suited to complex graph relationships (Buyukcakir et al., 9 Sep 2025).

A CGAT block is defined by

$$h_i^{(l+1)} = \text{LayerNorm}\left( {h'}_i^{(l)} + \sigma_2 \boldsymbol{W}_2^{(l)} \cdot {h'}_i^{(l)} \right),$$

$G(\mathcal{V}, \mathcal{E}),$0

Each block contains one GATv2 convolution, one linear layer, residual connections, LayerNorm, a nonlinearity typically GeLU, multi-head attention with 8 heads, and head merging by max-pooling (Buyukcakir et al., 9 Sep 2025). The authors note that this resembles a Transformer layer, but with graph attention instead of dot-product attention.

The CLS node is connected to all mesh nodes and is the sole input to the final classifier. Its functions are explicitly enumerated in the paper: it acts as a global information aggregator, receives information from all nodes, serves as the sole input to the classifier, and provides a single locus from which explanation can be extracted (Buyukcakir et al., 9 Sep 2025). This is the defining difference between CGAT and ordinary graph-attention models that aggregate graph information more diffusely.

The paper also studies two edge configurations for the CLS node: directed edges $G(\mathcal{V}, \mathcal{E}),$1 and undirected edges $G(\mathcal{V}, \mathcal{E}),$2. Directed edges let information flow from mesh nodes to CLS while preventing CLS from influencing node embeddings. Undirected edges allow bidirectional interaction and effectively make every node close to every other node through CLS. The paper argues that this can create a problematic global 2-hop neighborhood effect and make attention maps less intuitive in deeper models (Buyukcakir et al., 9 Sep 2025).

4. Attention rollout and explanation of model decisions

CGAT’s explanation mechanism is based on attention rollout across multiple graph-attention layers. The rollout equations are

$G(\mathcal{V}, \mathcal{E}),$3

$G(\mathcal{V}, \mathcal{E}),$4

Here $G(\mathcal{V}, \mathcal{E}),$5 is the attention matrix at layer $G(\mathcal{V}, \mathcal{E}),$6, $G(\mathcal{V}, \mathcal{E}),$7 represents residual connections, and $G(\mathcal{V}, \mathcal{E}),$8 accumulates attention through the network (Buyukcakir et al., 9 Sep 2025). The visualization focuses on the values corresponding to

$G(\mathcal{V}, \mathcal{E}),$9

which are used to show each node’s contribution to the final decision.

The qualitative claim of the paper is that the resulting attention maps align with clinical understanding of tooth development. Early stages attend to the crown and central occlusal region, mid stages shift attention toward the furcation and beginning roots, and late stages concentrate more strongly on roots and root apices (Buyukcakir et al., 9 Sep 2025). The paper presents this as the expected progression in Demirjian staging.

The comparison between directed and undirected CLS edges is central to the interpretability analysis. Directed edges produced attention maps that remained more distributed across multiple relevant anatomical regions, avoided collapsing attention into a single region, and better matched human reasoning. Undirected edges, especially in deeper models, tended to produce more concentrated and sometimes less interpretable attention because the CLS node influenced all nodes back through the graph (Buyukcakir et al., 9 Sep 2025).

This interpretability claim is specific. The paper does not merely state that attention exists; it argues that the architectural constraint on information flow determines whether attention maps are anatomically meaningful. A plausible implication is that, within CGAT, explanation quality is an emergent property of graph topology and message-passing directionality rather than a post hoc add-on.

5. Training protocol and empirical results

The reported dataset consists of 138 patients, 67 males and 71 females, providing 528 third molars with stages D to H (Buyukcakir et al., 9 Sep 2025). The class distribution is highly imbalanced: stage D has only 18 samples, whereas stage G has 351. The train/validation/test split is 80%/5%/15%.

Training uses cross-entropy loss, the Adam optimizer, an initial learning rate of 0.001, learning-rate reduction by a factor of 0.5 if validation loss plateaus for 5 epochs, 150 total epochs, and dropout with $751 \pm 1$0 on the CLS embedding (Buyukcakir et al., 9 Sep 2025). To handle class imbalance, the paper uses weighted sampling so that each batch is class-balanced. Models with 1 to 15 CGAT blocks were evaluated, all with 8 attention heads.

The primary metric is weighted F1 score, selected because of class imbalance; the paper also reports weighted precision, weighted recall, and MAE in comparison experiments (Buyukcakir et al., 9 Sep 2025). The best-performing configuration used both curvature and distance-to-centroid features and achieved about weighted F1 $751 \pm 1$1 and MAE $751 \pm 1$2. This is compared with the strongest baseline GAT at F1 $751 \pm 1$3 and MAE $751 \pm 1$4. The paper further states that CGAT outperformed PointNet++, MeshCNN, GCN, and GAT, while remaining computationally efficient, with inference times comparable to GAT and faster than PointNet++ and MeshCNN (Buyukcakir et al., 9 Sep 2025).

Depth was analyzed separately from feature choice. The paper reports no severe over-smoothing in the tested range; deeper models often improved or stabilized performance, and a practical recommended depth was about 6–12 blocks depending on feature set. For directed CLS edges, performance was relatively stable across depth (Buyukcakir et al., 9 Sep 2025). This suggests that the CLS-based aggregation did not, within the tested regime, impose the kind of degradation often associated with deeper message passing.

6. Terminological distinctions, related formulations, and common confusions

The term “CGAT” is not used uniformly across the literature, and two nearby usages should be distinguished from the Class Node Graph Attention Network of the 3D-shape paper.

First, there is a conceptual but non-nominal overlap with class-aware attention in node classification. The method proposed in “Supervised Attention Using Homophily in Graph Neural Networks” does not introduce a new class-node architecture; instead, it adds an auxiliary supervision loss to a standard attention-based GNN so that nodes learn to assign higher attention to same-class neighbors and lower attention to different-class neighbors (Chatzianastasis et al., 2023). Its backbone is GATv2, and the combined objective is

$751 \pm 1$5

where $751 \pm 1$6 is the node-classification cross-entropy and $751 \pm 1$7 is a binary cross-entropy over training edges derived from node labels. The resulting variant is termed HS-GATv2 rather than CGAT. The paper explicitly frames this as class-aware attention, but not as a model with explicit class nodes or a separate class-node graph (Chatzianastasis et al., 2023). Accordingly, if “CGAT” is interpreted broadly as attention guided by class membership, that work is adjacent in spirit; if “CGAT” is taken to mean an architecture with a dedicated class or CLS node, it is not the same object.

Second, “CGAT” can be confused with a contextualised graph attention model for relation extraction. “Contextualised Graph Attention for Improved Relation Extraction” proposes a contextualized GAT over three dependency-derived sub-graphs, with BERT-based token encodings, BiLSTM contextualisation, and edge-aware attention that concatenates edge features into the attention computation (Mandya et al., 2020). Its best model, $751 \pm 1$8, achieves an F1-score of 86.30 on SemEval-2010 Task 8. However, the “C” in that formulation stands for contextualised, not class-node, and the architecture contains neither class nodes nor class prototypes (Mandya et al., 2020).

These distinctions matter because they separate three different mechanisms that can all appear “class-aware” or “CGAT-like” at a superficial level: a global CLS-node architecture for graph-level classification and explanation (Buyukcakir et al., 9 Sep 2025), label-supervised attention that biases node-to-node aggregation toward homophily (Chatzianastasis et al., 2023), and contextualised graph attention for relation extraction with edge features and multiple sub-graphs (Mandya et al., 2020). The shared reliance on attention does not make them architecturally equivalent.

7. Significance and limitations

Within the reported study, CGAT is significant because it combines three properties that are often treated separately in graph-based 3D learning: competitive classification performance, computational efficiency, and interpretable attention maps (Buyukcakir et al., 9 Sep 2025). The paper emphasizes that this is particularly relevant in high-stakes settings where trust, validation, and accountability are essential.

The interpretability claim, however, is specific to the chosen setup. The strongest reported results and the most comprehensive attention visualizations were obtained when using both local mean curvature and distance to centroid as node features, together with directed edges to the CLS node (Buyukcakir et al., 9 Sep 2025). This suggests that CGAT’s explanatory behavior depends materially on graph construction, feature definition, and edge directionality, rather than arising automatically from the presence of attention.

The paper also presents CGAT as broadly applicable to graph-level classification, graph-level regression, other mesh-based shape analysis tasks, and graph data in general, including domains such as molecular analysis or other structured medical imaging problems (Buyukcakir et al., 9 Sep 2025). This is a statement of proposed scope rather than a cross-domain empirical validation. A plausible implication is that the central transferable idea is the use of a dedicated global node as both prediction bottleneck and explanation interface.

In that sense, Class Node Graph Attention Network denotes a specific synthesis: graph attention with a learnable global CLS node, classification from that node alone, and explanation through attention rollout into that node. Its contribution is not simply the use of attention on graphs, but the coupling of global aggregation and interpretable attribution within a single graph-level architecture (Buyukcakir et al., 9 Sep 2025).