TrafficKAN-GCN: Adaptive Graph Learning

• TrafficKAN-GCN is an adaptive framework that uses dynamic graph construction to overcome the limitations of static graph models in both image recognition and neuroimaging.
• The architecture integrates a semantic attention module for image-specific graph generation and multi-level connectome fusion for detailed hierarchical analysis.
• Experimental results indicate that TrafficKAN-GCN consistently outperforms baseline methods, demonstrating improvements in metrics like mAP and AUC across diverse datasets.

ADD-GCN encompasses two distinct, high-impact Graph Convolutional Network (GCN) architectures: (1) the Attention-Driven Dynamic Graph Convolutional Network for multi-label image recognition (Ye et al., 2020), and (2) the Multi-Level Generated Connectome GCN for Alzheimer's Disease analysis (Zhu et al., 2024). Both leverage dynamic or adaptive graph construction to encode subject/image-specific relationships, advancing state-of-the-art performance in their respective domains.

1. Foundations and Core Principles

ADD-GCN architectures are predicated on modeling highly adaptive, context-specific graph structures. Standard GCNs rely on static adjacency matrices, often based on global statistics such as co-occurrence or correlation; ADD-GCNs circumvent the limitations of static edges by learning or generating graph structures dynamically for each input (image or subject). In both image recognition and neuroimaging contexts, this enables richer, content-sensitive dependency modeling and robust generalization to the test domain.

The multi-label image recognition variant utilizes a semantic attention mechanism to generate category-specific representations and an image-dependent graph that reflects observed label correlations in each image (Ye et al., 2020). The neuroimaging variant generates multi-scale functional connectomes using hierarchical spatio-temporal feature extraction, then encodes these graphs independently through dedicated GCN modules, fusing the resulting embeddings for disease classification (Zhu et al., 2024).

2. Model Architectures

Multi-Label Image Recognition (ADD-GCN (Ye et al., 2020))

• Backbone: Images processed by ResNet-101, producing feature map $X \in \mathbb{R}^{H \times W \times D}$.
• Semantic Attention Module (SAM): Computes $C$ attention maps $M = [m_1, ..., m_C]$ via 1x1 convolution and classifier; each attended vector $v_c$ ($c=1...C$) is a spatially pooled, content-aware class vector, stacked into $V \in \mathbb{R}^{C \times D'}$.
• Dynamic GCN (D-GCN):
  • Static GCN: Global label graph $A_s$.
  • Dynamic GCN: Constructs image-specific adjacency $A_d$ from combined nodewise descriptors $h'_i = [h_i; h_g]$; propagates $Z = \mathrm{LeakyReLU}(A_d \cdot H \cdot W_d)$.
  • Output scores from GCN-refined descriptors ($s_r^c$) and SAM auxiliary scores ($s_m^c$) are averaged; sigmoid output thresholded at $0.5$ yields label predictions.

Multi-Level Connectome GCN for AD (MLC-GCN (Zhu et al., 2024))

• Multi-Graph Generation Block:
  • Each ROI time series $X \in \mathbb{R}^{n \times L}$ embedded via 1D-CNN + linear layer, then processed by $K$ stacked Spatio-Temporal Feature Extraction (STFE) blocks.
  • Each STFE produces features $h_i \in \mathbb{R}^{n \times d}$; connectome at each level $A^{(i)} = h_i h_i^T$.
• GCN Prediction Block:
  • For each graph (Pearson baseline $G_0$ and $K$ learned connectomes $\{G_1,...,G_K\}$), two-layer GCN encodes features and adjacency.
  • All graph embeddings $\{E_0,...,E_K\}$ concatenated; final prediction via MLP + softmax.
• Fusion: Only at classification stage, facilitating multi-scale integration of functional connectivity patterns.

3. Mathematical Frameworks and Learning

ADD-GCN (Ye et al., 2020)

• Attention Module:
  • $M = \sigma(\mathrm{Conv}_{cls}(X; W_m))$, $$v_c = \sum_{i=1}^H \sum_{j=1}^W m_c(i,j) \cdot x'_{i,j}$$.
• Dynamic Adjacency:
  • $A_d = \sigma(W_A \cdot [h'_1, ..., h'_C])$ where $h_g = \mathrm{GAP}(H)$ is the global node summary.
• Graph Propagation:
  • $Z = \mathrm{LeakyReLU}(A_d \cdot H \cdot W_d)$.
• Loss function: Binary cross-entropy over per-class scores.

MLC-GCN (Zhu et al., 2024)

• Input Embedding:
  • $$Z = \sigma(\mathrm{Flatten}(\mathrm{Conv}(X)) W) + \mathrm{PE}$$, with sinusoidal positional encoding.
• STFE Block: Dual-path – spatial Transformer encoder, temporal (trend/seasonal) DLinear + MLP.
• Connectome Creation: $A^{(i)} = h_i h_i^T$ at each hierarchy level.
• Graph Convolution:
  • $\widehat{A}^{(i)} = A^{(i)} + I$
  • $$h_{j+1}^{(i)} = \sigma(\widehat{A}^{(i)} h_{j}^{(i)} W^{(i)}_j)$$
  • Embeddings $E_i$ fused for classification.
• Losses: Categorical cross-entropy and intra-group regularization (encourages graph similarity within diagnostic labels).

4. Implementation Details and Training Protocols

ADD-GCN (Ye et al., 2020)

• Backbone: ResNet-101 (ImageNet pretraining), $D' = 1024$, $D_1 = 1024$, $D_2 = 512$.
• Nonlinearity: LeakyReLU ($\text{slope}=0.2$), sigmoid for attention/adjacency.
• Augmentation: Random resized crop, horizontal flip.
• Optimization: SGD, momentum $0.9$, weight decay $1 \times 10^{-4}$.
• Learning rates: Backbone $0.05$, SAM/D-GCN $0.5$; 50 epochs, scheduled decay.

MLC-GCN (Zhu et al., 2024)

• Preprocessing: Brainnetome toolkit, parcellation into $n=273$ ROIs, standard nuisance removal.
• Hyperparameters: AdamW, LR $0.001$, weight decay $0.001$, dropout $0.2$, 300 epochs, Mixup augmentation.
• STFE depth: $K \in \{6,12,24\}$, embedding $l=64$.
• Cross-validation: 5-fold stratified.
• Datasets: ADNI (643 scans), OASIS-3 (900 scans).

5. Experimental Results and Comparative Performance

ADD-GCN (Ye et al., 2020)

Dataset	ADD-GCN mAP	Prior SOTA mAP	Gain
MS-COCO (80cls)	85.2%	83.8 (SSGRL)	+1.4%
VOC2007 (20cls)	96.0%	95.0 (SSGRL)	+1.0%
VOC2012 (20cls)	95.5%	94.8 (SSGRL)	+0.7%

ADD-GCN consistently improves over SSGRL and ML-GCN baselines.

MLC-GCN (Zhu et al., 2024)

Task	MLC-GCN Acc	Baseline Acc	MLC-GCN AUC	Baseline AUC
NC vs. AD (ADNI, K=24)	95.74%	~93.4% (DABNet/LG-GNN)	97.76	~95.1
NC/MCI/AD (OASIS-3, K=24)	90.56%	89.3% (LG-GNN)	94.36	94.1

Ablation studies confirm necessity of both temporal/spatial branches and multi-level fusion; omitting these reduces accuracy and AUC by 1–3%.

6. Explainability, Limitations, and Future Research

MLC-GCN (Zhu et al., 2024) demonstrates sparser, more clinically focused connectomes, with salient edges in prefrontal and temporal lobes consistent with AD pathophysiology. Node rankings match established clinical findings (top regions include SFG/MFG/IFG, PCL, STG/MTG).

Limitations include reliance on vanilla 2-layer GCNs; the application of more expressive GNN variants (e.g., GAT, InceptionGCN) is suggested as a potential improvement. Current instantiations are limited to functional rs-fMRI data and validated only on AD, though extension to multimodal and wider clinical domains is anticipated.

7. Context, Impact, and Outlook

ADD-GCN architectures mark a progression from static graph modeling to input-specific graph construction in deep learning frameworks. For multi-label image classification, dynamic reasoning over content-aware category relationships yields state-of-the-art accuracy and robust generalization, mitigating biases inherent in training-set co-occurrence. For neuroimaging-based diagnosis, hierarchical multi-connectome integration facilitates both improved prediction and interpretable biomarker extraction.

A plausible implication is the generalization of ADD-GCN design patterns to heterogeneous, structured domains requiring adaptive relational inference, spanning vision, biomedical, and real-world sensor data. Advances in graph encoder architectures and graph construction mechanisms remain ongoing research directions.