CogGNN: Diverse Graph Neural Network Approaches

• CogGNN is a family label for distinct GNN formulations that regulate graph information flow using mechanisms like commute-time geometry, stochastic action selection, or cognitive integration.
• The directed commute-time variant uses precomputed, mutual random-walk reachability measures to weight neighbor aggregation, enhancing node classification on asymmetric graphs.
• The cooperative and connectomics models adaptively rewire communication—via sampled node actions or cognitive embeddings—to address diverse tasks from heterogeneous graph learning to generative connectomics.

CogGNN is a label applied to multiple, technically distinct graph neural network formulations rather than a single canonical architecture. In the literature represented here, it denotes three separate lines of work: a directed-graph model based on commute-time-weighted aggregation in "Commute Graph Neural Networks" (Zhuo et al., 2024), a dynamic communication framework in "Cooperative Graph Neural Networks" (Finkelshtein et al., 2023), and a cognition-aware generative framework for brain networks in "CogGNN: Cognitive Graph Neural Networks in Generative Connectomics" (Soussia et al., 13 Sep 2025). These uses share a concern with controlling graph information flow, but they differ in graph domain, objective, and mathematical machinery. A plausible implication is that the term is best treated as a family label whose meaning depends on context.

1. Terminological scope and variant structure

The three usages of CogGNN can be organized by the mechanism through which they alter or regulate graph computation (Zhuo et al., 2024, Finkelshtein et al., 2023, Soussia et al., 13 Sep 2025).

Usage of the label	Defining mechanism	Primary setting
Commute Graph Neural Networks	Commute-time-weighted, direction-aware message passing	Node classification on directed graphs
Cooperative Graph Neural Networks	Node-wise action selection over listening and broadcasting	General graph learning on synthetic, node, graph, and long-range tasks
Cognitive Graph Neural Networks in Generative Connectomics	Co-optimization of graph generation and visual-memory-based cognition	Multi-view brain network generation and CBT learning

The terminological overlap can obscure substantive differences. The commute-time model is rooted in Markov-chain geometry on digraphs. The cooperative model recasts message passing as a stochastic strategy-selection process in which nodes act as players. The connectomics model is not primarily a message-passing innovation for generic graph benchmarks; instead, it is a generative framework for connectional brain templates (CBTs) that couples a GNN backbone with an Echo State Network (ESN). Accordingly, any reference to “CogGNN” requires disambiguation by paper, objective, or application domain.

2. Commute-time CogGNN for directed graphs

In "Commute Graph Neural Networks" (Zhuo et al., 2024), the central problem is that standard GNNs on directed graphs capture unidirectional relations but fail to encode mutual path dependencies produced by asymmetric shortest paths and random-walk reachability. The model addresses this by integrating node-wise commute time into neighborhood aggregation. In this formulation, commute time between nodes $v_i$ and $v_j$ is

$c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$

where hitting times are direction-dependent. Small commute time indicates strong mutual proximity, whereas large commute time indicates weak mutual reachability.

The construction begins from a directed graph $G=(V,E,X)$ with adjacency $A$, out-degree matrix $D$, and transition matrix $P=D^{-1}A$. The method generalizes the Laplacian to directed graphs through a divergence over out-neighbors: $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$ To reflect global structural importance, it introduces the Perron-weighted digraph Laplacian

$$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$

where $\pi$ is the stationary distribution. The approach assumes irreducibility and aperiodicity. To ensure these conditions while preserving sparsity, it performs similarity-based rewiring: it computes an anchor feature $v_j$0, orders nodes by cosine similarity $v_j$1, connects them sequentially with bidirectional edges to form a strongly connected graph $v_j$2, and adds those edges and self-loops to the original graph. The rewired digraph $v_j$3 adds at most two extra edges per node.

Commute time is computed through the fundamental matrix

$v_j$4

but direct inversion is impractical. The paper derives a sparse formulation based on

$v_j$5

together with the lemma

$v_j$6

where $v_j$7 is the Moore–Penrose pseudoinverse. This yields explicit expressions for hitting and commute times from $v_j$8. Because $v_j$9 is sparse, the method approximates $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$0 with randomized truncated SVD: $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$1 at cost $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$2, with $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$3 typically approximately $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$4.

The resulting commute matrix $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$5 is converted to a proximity matrix by exponential damping,

$c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$6

then restricted to actual edges: $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$7 After row-wise max-normalization, these edge weights directly modulate neighbor contributions. A concrete layer instantiation follows the DirGNN pattern: $c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$8

$c(v_i,v_j)=h(v_i,v_j)+h(v_j,v_i),$9

$G=(V,E,X)$0

The weights are static across layers: they are precomputed once per graph and are not updated during message passing.

Empirically, the model is evaluated on eight directed benchmarks: Squirrel, Chameleon, Citeseer, Cora-ML, AM-Photo, Snap-Patents, Roman-Empire, and Arxiv-Year. It achieves state-of-the-art accuracy on 6 of 8 datasets and comparable performance on the remaining two, with representative accuracies of 77.61% on Squirrel, 79.54% on Chameleon, 70.27% on Citeseer, 77.06% on Cora-ML, 90.41% on AM-Photo, 72.33% on Snap-Patents, 92.89% on Roman-Empire, and 66.97% on Arxiv-Year. The paper also reports that the method improves especially on heterophilic graphs, that the similarity-based rewiring is both faster and more accurate than a dense Personalized PageRank alternative, and that symmetrizing the graph degrades performance. Per-layer message passing has complexity $G=(V,E,X)$1, overall time is $G=(V,E,X)$2, and experiments include graphs up to approximately 3 million nodes and approximately 14 million edges.

3. Cooperative CogGNN as dynamic communication control

In "Cooperative Graph Neural Networks" (Finkelshtein et al., 2023), CogGNN denotes a framework in which each node is a player that chooses one of four actions at every layer: Listen ($G=(V,E,X)$3), Broadcast ($G=(V,E,X)$4), Listen & Broadcast ($G=(V,E,X)$5, the standard case), or Isolate ($G=(V,E,X)$6). Standard message passing is recovered when every node chooses $G=(V,E,X)$7. The framework comprises two cooperating networks: an action network $G=(V,E,X)$8, which outputs action probabilities and samples actions using the straight-through Gumbel–Softmax estimator, and an environment network $G=(V,E,X)$9, which performs the actual state updates subject to the sampled communication pattern.

The baseline MPNN formulation is

$A$0

Co-GNN replaces fixed communication with a sampled, layer-specific directed computation graph. For each node,

$A$1

The induced directed edge set is

$A$2

and the environment update is

$A$3

$A$4

The same mechanism admits a gating reparameterization. If $A$5 and $A$6 denote listening and broadcasting gates, then

$A$7

$A$8

Action sampling is performed through the straight-through Gumbel–Softmax, with a learned inverse temperature

$A$9

The model is trained end-to-end on task loss, without requiring explicit gate regularization, although the formulation supports such terms.

Theoretical claims occupy a central place. The paper states that Co-GNN can probabilistically distinguish some graph pairs indistinguishable by 1-WL: $D$0 It also gives a long-range approximation theorem: $D$1 for compact differentiable $D$2, connected $D$3, a target $D$4, and sources $D$5. The discussion of over-squashing uses the standard MPNN bound

$D$6

and argues that Co-GNN mitigates the bottleneck by dynamically rewiring computation through action-selected directed paths.

Empirically, the framework is tested on synthetic, node-classification, graph-classification, and long-range benchmarks. On RootNeighbors regression, where the target is the average feature of root neighbors of degree 6, baseline MAE values are Random 0.474, GCN 0.468, GAT 0.442, SumGNN 0.370, SAGE 0.336, and MeanGNN 0.329, while Co-GNN (Sum action, Mean env) achieves 0.079. Action accuracy in a single-layer setup reaches 99.71% for Sum action/Mean env, 99.55% for Sum action/Sum env, and 57.20% for Mean action/Mean env. On heterophilous node classification, reported best Co-GNN results include roman-empire 91.57 ± 0.32, amazon-ratings 54.17 ± 0.37, minesweeper 97.31 ± 0.41, tolokers 84.45 ± 1.17, and questions 80.02 ± 0.86, with average improvement over baselines approximately +2.23%. On graph classification, it reports IMDB-B 72.2 ± 4.1, IMDB-M 49.9 ± 4.5, REDDIT-B 90.4 ± 1.9, NCI1 80.6 ± 1.1, and ENZYMES 68.3 ± 5.7. On the LRGB Peptides-func benchmark, Co-GNN (Sum action, Mean env) attains 0.6990 ± 0.0093 AP. On the Cycles expressivity dataset, it reaches 100% train accuracy while corresponding classical MPNNs achieve 50%. Runtime scales linearly relative to GCN with $D$7, and asymptotic complexity remains comparable to the base GCN when the action network is shallow.

4. CogGNN in generative connectomics

In "CogGNN: Cognitive Graph Neural Networks in Generative Connectomics" (Soussia et al., 13 Sep 2025), the term refers to a cognition-aware generative framework for multi-view brain graphs. The motivation is that prior GNN-based generative methods in connectomics focus on structural and topological fidelity while neglecting cognitive traits. CogGNN is introduced as the first cognition-aware generative GNN framework that explicitly incorporates cognitive capacities, demonstrated with visual memory, into graph generation. The reported instantiation, Vis-CogGNN, is designed for learning cognitively enhanced connectional brain templates.

The formal setting uses multi-view brain networks $D$8 with adjacency matrices $D$9, shared node features $P=D^{-1}A$0, and subject tensors $P=D^{-1}A$1. A generic cognition-aware generator is written as

$P=D^{-1}A$2

where $P=D^{-1}A$3 is a cognitive embedding and $P=D^{-1}A$4 captures population-level structure. For CBT learning, the model instead produces a single template $P=D^{-1}A$5.

Vis-CogGNN combines a GNN backbone with reservoir computing. The backbone is the Deep Graph Normalizer (DGN), implemented with edge-conditioned convolution (ECC). At ROI $P=D^{-1}A$6 and layer $P=D^{-1}A$7,

$P=D^{-1}A$8

with ReLU nonlinearity. The initial CBT is then formed by pairwise $P=D^{-1}A$9 distances in the final embedding space: $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$0

Cognitive integration is performed through an Echo State Network. Visual inputs $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$1, instantiated with MNIST digits resized to $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$2, are encoded as vectors $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$3. The reservoir topology is derived from the current CBT: $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$4 where $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$5 is the spectral radius. The state update is

$$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$6

and the readout is

$$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$7

The paper also describes an optional external memory bank with key–value retrieval and FiLM-style fusion, but notes that this external memory module is not instantiated in the reported experiments.

Training alternates between structural centeredness and visual-memory-based cognition. The CBT-centeredness loss is

$$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$8

with $$(Ts)_i=\sum_{j\in N_i^{out}}P_{ij}(s_i-s_j)=((D^{-1}-P)s)_i, \qquad T=D^{-1}-P.$$9 normalizing views by average weights. The cognitive loss is the delayed image reconstruction objective

$$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$0

The co-optimization objective is

$$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$1

followed by element-wise median aggregation of subject-specific CBTs to form a symmetric, non-negative population CBT $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$2.

The evaluation uses AD/LMCI and ASD/NC cohorts. AD/LMCI comprises 77 subjects, with 41 AD and 36 LMCI, and four morphological views per hemisphere derived from T1-weighted MRI with 35 ROIs under the Desikan–Killiany atlas. ASD/NC comprises 310 subjects, with 155 ASD and 155 NC, and six morphological views per hemisphere. Metrics include centeredness via mean Frobenius distance, Visual Memory Capacity (Vis-MC), topological measures such as information centrality, Laplacian centrality, and node strength, and SVM-based classification using ACC, SS, SP, AUC, and F1. Reported results state that Vis-CogGNN achieves lower Frobenius distances than DGN in 75% of cases, significantly outperforms DGN in visual memory capacity in 6/8 datasets at $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$3, and substantially improves discriminative performance. Examples include AD/LMCI RH with ACC 74% versus 51.67%, AUC 87.68 versus 52.47, and SP 100% versus 64.07; AD/LMCI LH with ACC 66% versus 55.32% and SS 100% versus 80.20; ASD/NC LH with ACC 68% versus 52.49%, AUC 84.32 versus 54.44, and SS 100% versus 34.58; and ASD/NC RH with ACC 68% versus 51.39%, AUC 84.32 versus 55.93, and SS 100% versus 22.33. The implementation uses PyTorch Geometric for graph learning and echoes for ESNs, with 3 ECC layers, spectral radius $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$4, leakage $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$5, input scaling $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$6, 5-fold cross-validation, Adam, 500 epochs, and early stopping.

5. Comparative methodological axes

Across the three usages, the main commonality is selective modulation of information flow, but the operational locus of that modulation differs materially (Zhuo et al., 2024, Finkelshtein et al., 2023, Soussia et al., 13 Sep 2025).

For commute-time CogGNN, selectivity is encoded in static, precomputed edge weights derived from mutual random-walk reachability on a directed graph. Directionality is explicit in both the underlying transition matrix and the masking of proximities by $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$7 and $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$8. The method does not learn the weighting scheme per layer; instead, it learns feature transforms while retaining graph-level commute priors throughout the network. This makes the model structurally biased toward mutual asymmetric proximity.

For cooperative CogGNN, selectivity is endogenous and stochastic. Each node chooses a communication role at each layer, so the computation graph becomes a learned directed subgraph of the original graph. The induced topology is layer-specific and task-adaptive. The model therefore separates the input graph from the computation graph more aggressively than standard attention-based MPNNs, because it can inhibit communication completely through $$\Pi=\mathrm{diag}(\pi), \qquad \pi^TP=\pi^T, \qquad \mathcal{T}=\Pi(D^{-1}-P),$$9 or enforce asymmetry through $\pi$0 and $\pi$1.

For connectomics CogGNN, selectivity is not primarily about local message routing in a benchmark graph. Instead, it arises from alternating optimization between structural template estimation and cognitive alignment. The GNN backbone shapes a CBT; the CBT then becomes the ESN reservoir topology; the reservoir’s delayed-recall performance feeds back into template learning. This introduces a different notion of graph computation, in which the graph is both the object being generated and the substrate for memory dynamics.

These differences also separate the three models by task family. Commute-time CogGNN is a semi-supervised node-classification architecture on digraphs. Cooperative CogGNN is a general framework applicable to node, graph, and long-range tasks on undirected input graphs. Connectomics CogGNN is a generative, multi-view neuroscience model aimed at CBT learning and cognitively grounded biomarker construction. A plausible implication is that direct empirical comparisons under a single benchmark protocol would be poorly posed, because the models solve different formal problems.

6. Misconceptions, limitations, and open directions

A common misconception is that CogGNN names a single research program. The evidence here supports the opposite: the label spans at least three non-equivalent frameworks, and only one of them is explicitly “cognitive” in the connectomics sense (Zhuo et al., 2024, Finkelshtein et al., 2023, Soussia et al., 13 Sep 2025). Another misconception is that all three models are primarily about attention or rewiring. In fact, one relies on commute-time geometry, one on stochastic action selection, and one on cognition-aware generative co-optimization.

Each formulation also carries distinct limitations. The commute-time model depends on similarity-based rewiring that is explicitly described as heuristic, uses static precomputed weights, and may offer limited gains in acyclic or inherently one-way interaction domains such as citation-like DAGs where mutual paths are structurally absent by design. The cooperative model relies on stochastic sampling, so its expressivity gains are probabilistic; the paper notes potential instability, possible degradation on homophilous graphs if communication becomes too sparse, added overhead for large graphs, and a lack of explicit theoretical convergence guarantees for the stochastic gated dynamics. The connectomics model depends on visual input as the cognitive channel, is evaluated on morphological networks with limited sample sizes, leaves interpretation of reservoir dynamics as an open issue, and does not instantiate the optional external memory bank in reported experiments.

The open directions likewise differ. For commute-time CogGNN, the paper points to more sophisticated, task-aware rewiring, adaptive or layer-dependent commute weights, and incremental approximations for very large or dynamic graphs. For cooperative CogGNN, natural extensions include temporal graphs, heterogeneous or multi-relational graphs, explicit communication budgets, entropy regularizers, hybrid attention–action networks, and tighter expressivity characterizations beyond 1-WL that do not rely on randomness alone. For connectomics CogGNN, proposed directions include extending cognition beyond visual memory to auditory memory or language comprehension, integrating stronger generative components such as diffusion, VAE, or GAN machinery, and theoretically analyzing how CBT-derived reservoir topology affects memory capacity and information flow. Taken together, these agendas suggest that “CogGNN” currently functions less as a settled architecture than as a naming locus for research on graph computation under richer notions of proximity, agency, or cognition.