Edge-Aware Graph Neural Networks

• Edge-aware GNNs are advanced graph neural network models that integrate rich edge semantics and attributes into the message passing process for improved representational power.
• They employ dual message passing, channel-wise decomposition, and attention mechanisms to jointly update node and edge embeddings in complex graph structures.
• Empirical evaluations demonstrate accuracy gains of up to 10.7 points and significant efficiency improvements, enabling effective deployment even on resource-constrained edge devices.

Edge-Aware Graph Neural Networks (Edge-aware GNNs) generalize classical graph neural models by explicitly incorporating the semantics, attributes, and topology of edges into the representation learning pipeline. Unlike standard approaches that treat edges as unstructured connection mechanisms or simple adjacency weights, edge-aware architectures encode, propagate, and leverage multidimensional edge features—ranging from learned embeddings, relation labels, channel weights, geometric attributes, and topological modalities (e.g., orientation, direction)—at all message-passing stages. This results in models with finer structural awareness, higher expressive power, improved interpretability, and stronger performance across node, edge, and graph-centric prediction tasks.

1. Architectural Foundations and Message Passing

Edge-aware GNNs are characterized by the integration of rich edge information into the message aggregation, feature updates, and optimization pipeline. Core architectural motifs include:

• Dual message-passing: Simultaneous update of node and edge embeddings (e.g., EGNAS (Cai et al., 2021)). At layer $l$, node $i$ aggregates via

$$h_i^{(l+1)} = \sum_{j \in \mathcal N(i)} \sum_{k=1}^{K_E}\alpha_k^E f_k^E(h_i^{(l)}, h_j^{(l)}, e_{ij}^{(l)})$$

and edge $(i,j)$ is updated by

$$e_{ij}^{(l+1)} = \sum_{k=1}^{K_F} \alpha_k^F f_k^F(h_i^{(l)}, h_j^{(l)}, e_{ij}^{(l)})$$

where $f^E_k$ and $f^F_k$ are candidate update functions parameterized by node/edge attributes.

• Channel-wise decomposition: Each edge is split or projected onto multiple structural or semantic channels, controlling multi-way message flow (EGD-GNN (Li et al., 2021), EdgeGFL (Zhuo et al., 4 Feb 2025)). For edge $(i,j)$ and channel $k$,

$$\alpha_{ij}^{(k)} = \frac{\exp(z_i^{(k)} \cdot z_j^{(k)})}{\sum_{k'} \exp(z_i^{(k')} \cdot z_j^{(k')})}$$

with $\sum_k \alpha_{ij}^{(k)} = 1$.

• Edge-aware attention mechanisms: Masking and gating of messages using edge attributes within an attention or gating formulation (Hybrid Supervised/Self-supervised GNN (Borzone et al., 21 Jan 2025), EdgeGFL (Zhuo et al., 4 Feb 2025)):

$$m^l_{u \leftarrow v} = \alpha^l_{uv} h^l_v + (1 - \alpha^l_{uv}) (h^l_v \odot r^l_{uv})$$

• Topology refinement via edge multiplicity: Usage of statistical generative models to infer latent edge strengths or multi-edges, then use the resulting weighted adjacency in deep GNN stacks (EEGNN (Liu et al., 2022)):

$$\hat{A}_{ij} = z_{ij}, \quad \hat{P} = \hat{D}^{-1/2} \hat{A} \hat{D}^{-1/2}$$

2. Expressive Power and Weisfeiler-Leman Extensions

Several edge-aware GNNs are motivated by limitations of standard message passing as formalized by the 1-dimensional Weisfeiler-Leman (1-WL) algorithm. Significant advances include:

• Explicit neighbor-edge structure: Models such as NEAR (Kim et al., 2019) and NC-GNN (Liu et al., 2022) augment node-wise updates with local edge configurations. For a node $v$ at layer $k$:

$$h_v^{(k+1)} = \mathrm{MLP}^{(k)}\left([h_v^{(k)} + h_{N_v}^{(k)}] \Vert h_{NE_v}^{(k)}\right)$$

where $h_{NE_v}$ aggregates $g(h_u, h_z)$ over $(u,z)$ edges among $v$’s neighbors.

• NC-1-WL and neural analogs: The NC-1-WL algorithm hashes not only neighbor multiset but also edges among neighbors, allowing differentiation of motifs (e.g., triangles vs. cycles) inaccessible to standard 1-WL and GIN.

$$h_v^{(\ell)} = \mathrm{MLP}_1^{(\ell)}((1+\epsilon^{(\ell)}) h_v^{(\ell-1)} + \sum_{u\in N(v)} h_u^{(\ell-1)} + \sum_{(u_1,u_2) \in E, u_1, u_2 \in N(v)} \mathrm{MLP}_2^{(\ell)}(h_{u_1}^{(\ell-1)} + h_{u_2}^{(\ell-1)}))$$

This structure yields provable expressiveness strictly between 1-WL and 3-WL (Liu et al., 2022).

• Gated-GIN and universal approximation: Integrates edge feature convolutions and GRU-style gating, subsuming the function classes of GIN and GG-NN, with full support for arbitrary edge attributes (Errica et al., 2020).

3. Edge Feature Construction, Propagation, and Channelization

Edge-aware GNNs employ sophisticated schemes for edge feature construction and evolution. Key techniques:

• Multidimensional edge embeddings: Rather than scalar adjacency, edge features $r^l_{ij} \in \mathbb{R}^{d_e^l}$ are jointly learned and projected into the appropriate node feature space (EdgeGFL (Zhuo et al., 4 Feb 2025)):

$r^l_{ij} = \delta(\hat{r}^l_{ij} W_r^l + b_r^l)$

• Multi-channel filter design: Edge embedding matrices are aggregated to synthesize $K$ different convolutional filters, each capturing a structural motif or relation (Zhuo et al., 4 Feb 2025):

$W^{l,(k)} = g_\phi(E^l)$

Edge-weighted adjacencies $A^{l,(k)}$ reflect channel-specific relations.

• Channel-aware attention and gating: Attention scores computed from node and edge features, integrated as residuals or gates in message passing (Zhuo et al., 4 Feb 2025).

4. Specialized Architectures for Edge-Level and Edge-Centric Tasks

Edge-aware GNNs have been tailored to support edge-centric prediction, topological signal modeling, and relational learning:

• Edge-centric supervised/self-supervised modeling: Hybrid models process node and edge features to predict relationships (e.g., protein-protein interactions, semantic similarity) via edge-aware attention, joint MLPs and permutation-invariant representations (Borzone et al., 21 Jan 2025).
• Edge-level topological GNNs: EIGN provides orientation equivariant/invariant shift operators that distinguish directed and undirected signals, leverage magnetic (complex-phase) propagation, and fuse equivariant/invariant channels for general-purpose edge-signal prediction (Fuchsgruber et al., 2024):

$$H_{\rm equ}^{(l)} = \sigma_{\rm equ}( L_{\rm equ}^{(q)} H_{\rm equ}^{(l-1)} W_{ee}^{(l)} + L_{\rm equ\leftarrow inv}^{(q)} H_{\rm inv}^{(l-1)} W_{ie}^{(l)} + H_{\rm equ}^{(l-1)} W_{e0}^{(l)} )$$

$$H_{\rm inv}^{(l)} = \sigma_{\rm inv}( L_{\rm inv}^{(q)} H_{\rm inv}^{(l-1)} W_{ii}^{(l)} + L_{\rm inv\leftarrow equ}^{(q)} H_{\rm equ}^{(l-1)} W_{ei}^{(l)} + H_{\rm inv}^{(l-1)} W_{i0}^{(l)} )$$

5. Edge-Aware Neural Architecture Search and Deployment on Edge Devices

Adaptive architecture search frameworks incorporate edge-awareness for performant, resource-constrained deployment:

• Fine-grained NAS with edge-featured search spaces: EGNAS discovers optimal node and edge update operators via differentiable bi-level optimization, mixing candidate atomic functions parameterized by learned edge embeddings. Search space topology incorporates rich feature dependence (node and edge DAGs) (Cai et al., 2021).
• Hardware- and edge device-aware GNNs: HGNAS introduces latency/memory predictors built from operation/function graphs processed by GCN+MLP, estimates peak memory per forward pass, and constrains NAS objectives to guarantee real-time deployment on edge platforms (RTX3080, Jetson TX2, Pi) (Zhou et al., 2024). The search stages decouple function and operation selection, yielding architectures with $>7$–$10\times$ speedup and $>40$–$80\%$ memory reduction.

6. Empirical Validation and Performance Gains

Edge-aware GNNs consistently demonstrate improved predictive performance and interpretability over conventional baselines:

• Node and graph classification: Gains of $+7$–$10.7$ points in accuracy for deep models with principled Bayesian edge modeling (EEGNN (Liu et al., 2022)); consistent improvements across benchmarks for edge aggregation and channelization (EGD-GNN (Li et al., 2021), NEAR (Kim et al., 2019), NC-GNN (Liu et al., 2022)).
• Edge-centric and relational tasks: MAE reductions and F1-score improvements in protein interaction, gene ontology, and compound similarity (Hybrid edge-aware GNN (Borzone et al., 21 Jan 2025)); up to $+5.4\%$ absolute Micro/Macro-F1 improvement for node classification on heterogeneous graphs with high-dimensional edge features (EdgeGFL (Zhuo et al., 4 Feb 2025)).
• Topological edge-level tasks: EIGN outperforms spectral and Hodge-based GNNs by up to $43.5\%$ RMSE reduction in traffic/circuit simulation, uniquely satisfying orientation equivariance/invariance and one-way directional constraints (Fuchsgruber et al., 2024).
• Device efficiency: HGNAS achieves $10.6\times$ speedup and $82.5\%$ memory reduction versus DGCNN with $<1\%$ accuracy loss (Zhou et al., 2024).

7. Theoretical Limitations, Scalability, and Practical Considerations

• Expressiveness–efficiency trade-offs: While channel-wise and edge-convolutional architectures achieve higher discriminative power (beyond 1-WL), they incur increased computation and memory—especially in dense graphs or models operating over multi-edge motifs (NC-GNN (Liu et al., 2022), Gated-GIN (Errica et al., 2020)).
• Parameterization and overfitting: Models with numerous gates, MLPs, or channel parameters (Gated-GIN, EdgeGFL) require regularization and ablation to avoid overfitting, especially on small datasets (Errica et al., 2020).
• Applicability to dynamic and heterogeneous graphs: Edge-aware designs generalize across tasks, including link prediction, heterogeneous graph learning, and graph-level pooling, provided appropriate projection and update mechanisms for edge attributes (Zhuo et al., 4 Feb 2025, Li et al., 2021).

Edge-aware Graph Neural Networks form a rich class of relational graph models that systematically incorporate edge semantics into representation learning pipelines. By designing multi-channel aggregators, adaptive edge embeddings, topologically-aware update rules, and attention/gating mechanisms, these models overcome expressiveness bottlenecks, enable edge-centric prediction, and support resource-efficient deployment—even on constrained edge devices. Empirical and theoretical advances continue to drive the development of edge-aware architectures for increasingly complex, heterogeneous, and dynamic graph domains.