Bidirectional & Stable GCN Architectures

• Bidirectional and stable GCN architectures are advanced neural networks that employ invertible mappings and dual filtering to robustly propagate information in both feature and topological domains.
• They integrate explicit bidirectional convolution mechanisms, leveraging bijective transforms, ADMM updates, and attention-based aggregation to denoise and refine graph signals.
• Empirical results reveal improved accuracy and stability in noisy environments, with applications in node classification, latent exemplar design, and text-rich network analysis.

Bidirectional and stable Graph Convolutional Network (GCN) architectures comprise a suite of advanced network designs characterized by robust information propagation in both feature and topological dimensions and rigorously controlled stability properties. These architectures address issues such as label efficiency, over-smoothing, feature denoising, and structural robustness in graph learning contexts by formalizing invertible/lipschitz mappings, bidirectional filtering, and heterogeneous message passing mechanisms.

1. High-Level Architectural Concepts

Bidirectional GCNs generalize standard GCN propagation by creating explicit two-way mappings between the ambient graph domain and learned latent spaces, or by smoothing both across node neighborhoods and feature correlations. In (Sahbi, 26 Nov 2025), the architecture is constructed as a bijective mapping $f : \mathbb{R}^p \to \mathbb{R}^p$, where the input graph signal is encoded into a latent space—enforced to be approximately Gaussian—and can be precisely reconstructed via the inverse $f^{-1}$. Layerwise invertibility is guaranteed by square weight matrices and bijective activations, often leaky-ReLU with tightly controlled slopes. This facilitates latent-space operations (such as exemplar search) with guaranteed correspondence to the data manifold.

Complementary approaches, such as BiGCN (Chen et al., 2021), extend GCNs via bidirectional low-pass filtering: convolution is performed both along the original graph structure and an additional learned feature graph, enabling simultaneous denoising of node-wise and feature-wise degradations.

In text-rich, multi-modal settings, BiTe-GCN (Jin et al., 2020) formalizes bidirectionality as explicit convolutions in both topological (document/word co-occurrence) and feature (textual content) spaces on bi-typed heterogeneous networks, balancing message diffusion via attention between four channels—topology (document-to-document, word-to-word) and cross-modal inclusion edges (document-to-word, word-to-document).

2. Mathematical Frameworks for Bidirectional Mapping

Bidirectional GCNs model both encoding and decoding flows through layerwise invertible transforms. For (Sahbi, 26 Nov 2025), the forward pass comprises:

$\phi^\ell = g_\ell(W_\ell^\top A\phi^{\ell-1})$

where $g_\ell$ is bijective, $W_\ell$ is invertible, and $A$ is fixed or pseudo-invertible. The entire network $f(\phi^0) = \phi^L$ is invertible; its inverse is constructed recursively as:

$$\phi^{\ell-1} = A^{-1} W_\ell^{-\top} g_\ell^{-1}(\phi^\ell)$$

This closed-form inversion supports reliable latent-to-ambient graph reconstruction.

BiGCN (Chen et al., 2021) models bidirectional low-pass filtering via minimization:

$$\min_Y \|Y - F\|_F^2 + \lambda_1 \textrm{tr}(Y^\top L_1 Y) + \lambda_2 \textrm{tr}(Y L_2 Y^\top)$$

yielding the Sylvester equation $(I + \lambda_1 L_1)Y + \lambda_2 Y L_2 = F$. ADMM splitting and Taylor expansion yields tractable updates at each layer, with explicit smoothing along both node and feature graphs.

In BiTe-GCN, the propagation equation generalizes to:

$$H^{(\ell+1)} = \sigma\left(\sum_{t \in T} \alpha_t \hat{A}_t H^{(\ell)} W_t^{(\ell)}\right)$$

where $T = \{\textrm{DD, WW, DW, WD}\}$ spans document and word graph adjacencies and inclusion maps, each weighted by $\alpha_t$.

3. Stability: Bi-Lipschitz Continuity and Regularization

Stability in bidirectional GCNs is formalized via bi-Lipschitz continuity. A network $f$ is bi-Lipschitz if both $f$ and $f^{-1}$ satisfy:

$$\|f(x)-f(y)\| \leq K \|x-y\| \quad , \quad \|f^{-1}(u)-f^{-1}(v)\| \leq M \|u-v\|$$

The stability product $KM$ is closely monitored. If all activations $g_\ell$ satisfy $l \leq |g_\ell'| \leq u$ and $\mathrm{cond}(W_\ell) \leq \kappa$, then $KM = (\kappa u / l)^{L-1}$. To enforce $KM \approx 1$, leaky-ReLU slopes $u \approx 0.99$, $l \approx 0.95$ and condition numbers $\kappa \approx 1$ are tightly regularized:

• Condition-number penalty: $$L_\text{total} = \mathrm{CE}(f) + \lambda\sum_\ell \|W_\ell\|_2 \|W_\ell^{-1}\|_2$$
• Orthonormal penalty: $$L_\text{total} = \mathrm{CE}(f) + \lambda \sum_\ell \|W_\ell^\top W_\ell - I\|_F$$
• Weight reparameterization: $W_\ell = \hat{W}_\ell + \delta I$, shifting the spectrum for stability.

BiGCN analyzes functional perturbation stability of spectral filters. For adjacency or feature graph drift $A \rightarrow A + \Delta A$, the propagated outputs remain stable under Kato’s matrix perturbation bounds or analogous row-wise analysis for feature Laplacians.

4. Layer Stack Design and Integration

Bidirectional and stable properties result from careful layerwise design. In (Sahbi, 26 Nov 2025), each layer comprises:

• Neighborhood aggregation via $A$
• Linear transformation $W_\ell^\top$
• Pointwise leaky-ReLU (slopes $[l, u]$)
• Optional single-head attention or skip connections
• No batch-norm or dropout; regularization via weight/activation constraints suffices

At inference time, the network propagates input features forward; for latent exemplar design, the closed-form inverse is called to map latent vectors back to input graphs.

BiGCN's layers (Chen et al., 2021) include:

• Fixed node Laplacian $L_1$
• Learnable feature Laplacian $L_2$
• ADMM updates $(p, K, \lambda_1, \lambda_2)$
• Linear transform $W^{(l)}$
• Dropout (0.5) and weight decay (5e-4) for generalization

BiTe-GCN (Jin et al., 2020) operates over bi-typed graphs, with each propagation step aggregating from four adjacency channels, with attention-based aggregation for each node type (documents and words).

5. Empirical Evaluation and Ablation Analysis

Bidirectional and stability-enforcing mechanisms demonstrate consistent empirical improvements across tasks and datasets.

• In (Sahbi, 26 Nov 2025), orthonormal regularization alone increased accuracy from 89.23% to 93.84% on SBU at 45% labels, with condition number dropping to 5.4 and FID to 10.2. Weight reparameterization (moderate $\delta$) further suppressed instability, but excessive rigidity or instability could arise if $\delta$ was not properly tuned. Latent-space exemplar design using the bijective mapping outperformed ambient design by 3–5% under label scarcity.
• BiGCN (Chen et al., 2021) surpassed baseline GCNs in node classification and link prediction under clean and noisy conditions; for instance, PubMed accuracy degraded from 80.0% to 69.1% with severe feature noise ($\sigma=0.9$) vs GCN’s 78.9% to 55.2%. Under structure mistakes, BiGCN lost only a few percent accuracy, while standard GCNs dropped >20%.
• BiTe-GCN (Jin et al., 2020) achieved higher accuracy than GCN, GAT, and other baselines on text-rich networks, e.g., 93.70% vs 88.15% (GCN) on Cora-Enrich. Deep stacks of 5–10 layers maintained accuracy within 1.5% of optimal, demonstrating effective prevention of over-smoothing.

Model	SBU (45% labels) Acc	Condition Number	FID Score
No regularizer	89.23%	Massive	Huge
OR only	93.84%	5.4	10.2
WR+OR	Higher stability	Lower	Lower

Data from (Sahbi, 26 Nov 2025); further extensive ablation tables contained therein.

6. Application Domains and Relevance

Bidirectional and stable GCNs find utility primarily in domains requiring reliable latent space manipulation, active learning with limited labels, robust action recognition, and text-rich relational reasoning. Concrete use cases include skeleton-based action recognition (Sahbi, 26 Nov 2025), node classification and link prediction with structural/feature noise (Chen et al., 2021), and text-rich citation analysis or e-commerce search (Jin et al., 2020).

Their bidirectionality enables data-driven exemplar and representation design in latent spaces with guaranteed invertible mapping, while stability bounds ensure that small latent perturbations correspond to small reconstructions in the input space. These properties are crucial in active learning pipelines, feature denoising, and the deployment of deep stacks with minimal over-smoothing.

7. Connections, Distinctions, and Practical Considerations

Bidirectional GCN architectures improve upon classical unidirectional designs (e.g., vanilla GCNs) by leveraging invertibility, latent-exemplar design, and cross-modal propagation. Stabilization via bi-Lipschitz constraints, orthonormality, and spectral conditioning distinguishes these networks for tasks sensitive to perturbations and label scarcity.

Implementation requires careful initialization and regularization of the weight matrices, selection of ADMM and stacking hyperparameters, and attention mechanisms for multi-channel information aggregation, as detailed in (Sahbi, 26 Nov 2025, Chen et al., 2021), and (Jin et al., 2020). There are scenario-specific considerations: for instance, weight reparameterization can become counterproductive if parameterized too aggressively, and proper balancing of topology-vs-feature convolutions is necessary in heterogeneous graphs.

A plausible implication is that future GCN designs will continue to unify invertibility, stability, and multi-channel propagation principles, enabling deeper, more robust graph architectures applicable to increasingly sparse and noisy domains.