Cognitive Graph Neural Networks

• Cognitive graph neural networks are graph‐based models that integrate cognitive variables like concept mastery and intelligence scores using adaptive relational inference and explicit reasoning.
• They employ diverse architectures—ranging from adaptive GNNs and neuro‐symbolic systems to connectome-based predictors—to enhance tasks in education and neuroimaging.
• Recent advancements focus on dynamic graph construction, uncertainty-aware edge differentiation, and multimodal longitudinal modeling to improve interpretability and scalability.

A cognitive graph neural network is a graph-based neural model in which graph construction, message passing, graph-level prediction, or explicit reasoning is tied to cognitive variables such as concept mastery, reasoning paths, cognitive task labels, intelligence scores, or trajectories of impairment. Across the literature, the label covers several related formulations: adaptive GNNs that jointly learn node features and graph structure, neuro-symbolic systems that combine graph encoders with explicit reasoning, and connectome-based architectures that map brain graphs to cognitive states or traits (Tang et al., 2019, Ding et al., 2019, Qu et al., 2021).

1. Conceptual scope and principal usages

In intelligent education, a cognitive graph neural network is a graph-based cognitive diagnosis model whose target variables are latent cognitive states. In the formulation of graph-based cognitive diagnosis, response logs define a student–exercise bipartite graph with student nodes, exercise nodes, and edges carrying correct or incorrect response semantics. The objective is to infer a student–concept proficiency matrix, and the model is termed “cognitive” because the target variables are cognitive states, the graph structure comes from cognitive behavior, and the prediction layer is a diagnosis function linking abilities, difficulties, and discriminations to response probabilities (Shao et al., 2024).

In connectomics, the term denotes models that map brain connectivity graphs derived from neuroimaging to cognitive outputs. The outputs may be discrete task labels, as in cognitive task classification from fMRI connectomes, or continuous traits such as fluid, crystallized, or total intelligence. In these formulations, nodes are brain regions or intrinsic connectivity networks, edges are structural or functional connections, and the model learns a mapping from network organization to cognition-related labels or scores (Maji et al., 31 Dec 2025, Thapaliya et al., 2023).

A further usage appears in explicit reasoning systems. CogQA defines a cognitive graph as a directed graph whose nodes are hop entities or answer spans and whose edges encode reasoning steps discovered during multi-hop question answering. The graph functions as an explicit working-memory structure coordinating an implicit extraction module (“System 1”) and an explicit reasoning module (“System 2”) (Ding et al., 2019).

Taken together, these usages suggest that “cognitive graph neural network” is best understood as a family of graph-based models in which representations and graph operations are organized around cognitive constructs rather than around graph prediction alone.

2. Adaptive relational inference and graph construction

One major line of work treats cognition-like behavior as adaptive relational inference. The Joint Learning Graph Convolutional Network (JLGCN) replaces a fixed adjacency with a layer-wise learned graph derived from current node features. At layer $l$, the edge weight between nodes $i$ and $j$ is defined by a Gaussian kernel over a Mahalanobis distance,

$$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$

with $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$, so graph learning becomes optimization of a low-dimensional matrix rather than a full PSD metric. The learned graph is normalized and combined with prior adjacency, and feature propagation proceeds as

$$\mathbf{F}_l = \mathbf{A}_l \mathbf{F}_{l-1} \mathbf{W}_l.$$

The training objective combines cross-entropy with a graph Laplacian regularizer, jointly optimizing representations and relational structure. On citation benchmarks, JLGCN reports 83.9% on Cora, 73.7% on Citeseer, and 79.7% on the 10,000-node Pubmed subset; on ModelNet40 it reports 90.8% overall accuracy and 87.2% mean class accuracy. The paper also reports that when the graph is empty, the model can still construct an effective adjacency from features alone (Tang et al., 2019).

A related mechanism appears in graph-based cognitive diagnosis under edge heterogeneity and uncertainty. ISG-CD treats correct and incorrect responses as distinct edge types, splitting the student–exercise graph into $X_1 = \{(s,e)\mid x_{se}=1\}$ and $X_0 = \{(s,e)\mid x_{se}=0\}$, and performs edge-type-specific message passing with separate transformations $W_1$ and $W_0$. To suppress uncertain edges arising from guessing or carelessness, the model introduces an Informative Edge Differentiation layer and selects reliable edges by confidence thresholds,

$i$0

The paper formulates this from an information bottleneck perspective, maximizing mutual information between the reliable graph and response logs while minimizing mutual information between the reliable graph and the original graph, and uses an alternating training strategy for the semantic-aware GNN and the edge differentiation layer (Shao et al., 2024).

These models operationalize “cognitive” primarily as continual revision of relational structure under evolving internal states, rather than as a fixed graph supplied in advance.

3. Neuro-symbolic reasoning and explicit cognitive structure

A second line of work makes the cognitive component explicit by coupling graph encoders with symbolic or quasi-symbolic reasoning. CogQA is organized around dual process theory. System 1 is a BERT-based implicit extraction module that reads a question, clue sentences, and an entity paragraph; System 2 is a GNN-based explicit reasoning module operating on a cognitive graph whose nodes are entities or answer candidates and whose edges encode reasoning steps. The graph is expanded iteratively, and the final answer is selected from answer nodes after graph reasoning. On HotpotQA fullwiki, the framework achieved a joint $i$1 score of 34.9, compared to 23.6 for the best competitor reported in the paper, while also exposing reasoning paths through the graph structure (Ding et al., 2019).

Neural-symbolic recommendation provides another formulation. GNNLR constructs an item–item graph from adjacent user interactions and uses a GCN to capture global implicit information,

$i$2

then converts user histories into propositional logic expressions and evaluates them with neural realizations of logical operators. The NOT and OR operators are parameterized as multilayer perceptrons,

$i$3

and recommendation is driven by similarity between a computed expression embedding and a benchmark truth vector. The model combines pairwise ranking loss with a logic-rule regularizer, thereby joining graph-derived implicit reasoning with local explicit logic reasoning (Chen et al., 2023).

Graph Reasoning Networks move the symbolic component into a differentiable satisfiability solver. A semi-learnable encoder combines a fixed canonicalized adjacency-string representation with a learned GNN embedding, and a SatNet-based reasoner learns logical rules over the resulting variables. On synthetic datasets involving connectivity and motif predicates, the adjacency-string version strongly outperforms the pure GNN baseline; on real-world datasets it is broadly comparable, with particularly strong behavior on topology-sensitive settings such as PROTEINS (Zopf et al., 2024).

A probabilistic precursor to these systems is the Contextual Graph Markov Model, which assigns each vertex a discrete hidden state $i$4, propagates context from preceding layers, learns by EM, and encodes a graph through layer-wise state frequencies. Although not labeled “cognitive” in the paper, its layered hidden-state semantics, explicit context diffusion, and graph fingerprints make it relevant to the same design space (Bacciu et al., 2018).

These systems suggest a distinct interpretation of cognitive GNNs: graph models in which message passing is supplemented by explicit reasoning objects—paths, clauses, latent states, or satisfiability constraints.

4. Connectome-based cognitive state and trait prediction

In brain-network applications, cognitive GNNs are typically connectome encoders for task decoding or trait prediction. SpectralBrainGNN defines each subject graph from task fMRI by using the Schaefer atlas with $i$5 ROIs, Pearson-correlation functional connectivity, and the normalized Laplacian

$i$6

The model performs exact graph Fourier transforms $i$7, applies learnable spectral filters $i$8, returns to the node domain, and aggregates node embeddings with an attention readout. On the HCPTask dataset, it reports 96.25% accuracy, 95.46% precision, 94.32% recall, and 95.58% $i$9, with task confusions concentrated in biologically plausible pairs such as Emotion/Gambling and Relational/Working Memory (Maji et al., 31 Dec 2025).

BrainRGIN predicts fluid, crystallized, and total intelligence from resting-state functional network connectivity. It introduces a clustering-based embedding in which each ROI-specific convolution weight is written as

$j$0

with $j$1 aligned to canonical brain subnetworks, and embeds this within a Graph Isomorphism Network-style update that also uses edge weights. The architecture combines TopK pooling with attention-based readout, and on ABCD reports best configurations of MSE 263 with correlation 0.23 for fluid intelligence, MSE 263.7 with correlation 0.30 for crystallized intelligence, and MSE 261 with correlation 0.31 for total composite scores. The middle frontal gyrus is repeatedly identified as important, with middle temporal gyrus and caudate nucleus also emerging for specific intelligence factors (Thapaliya et al., 2023).

MGCN addresses multi-modal cognitive ability prediction from task fMRI. It processes each subject’s graph with a two-layer GCN, fuses modality-specific graph embeddings through dense layers, and regularizes them by an ensemble manifold term

$j$2

where $j$3 is derived from within-modality and between-modality subject similarity matrices. On the Philadelphia Neurodevelopmental Cohort, for WRAT prediction, it reports RMSE $j$4 and MAE $j$5, outperforming single-modality GCNs, multi-task manifold baselines, and a multi-view GCN variant (Qu et al., 2021).

This branch of the literature defines the cognitive graph neural network as a graph-level predictor whose input graph is a connectome and whose output is a cognitive state, task, or trait.

5. Multimodal, generative, and longitudinal formulations

Recent work expands the concept beyond static graph-to-label prediction. MAGNet introduces a Transformer-style graph architecture for intelligence prediction from jointly modeled structural and functional connectomes. Structural information is derived from source-based morphometry, functional information from resting-state functional network connectivity, and the final hybrid graph is

$j$6

combining direct structural edges, direct functional edges, cross-modal connections, and multi-scale detour connectivity. Local edge-aware attention and global self-attention refine the hybrid graph, and training uses a joint loss

$j$7

with $j$8 and $j$9. On the ABCD cohort, MAGNet reports correlations of $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$0 for fluid intelligence, $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$1 for crystallized intelligence, and $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$2 for composite intelligence, while ablations show marked drops when multi-scale detour connectivity, cross-modal connections, or the structure–function consistency loss are removed (Mazumder et al., 31 Mar 2026).

CogGNN extends the field into generative connectomics. In Vis-CogGNN, a GNN generates subject-level connectional brain templates, and each generated template is used as the recurrent weight matrix of an Echo State Network. The reservoir state update is

$$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$3

with $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$4 instantiated from the generated brain graph. Training minimizes a structural objective together with a visual-memory cognitive loss over delayed image reconstruction, and evaluation uses Visual Memory Capacity,

$$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$5

The model is presented as a first “cognified” generative framework in which generated graphs are required to be both structurally meaningful and cognitively effective as dynamical reservoirs (Soussia et al., 13 Sep 2025).

Longitudinal connectome modeling appears in the history-aware GNN for Alzheimer’s disease progression. Each visit is encoded as a functional connectivity graph from rs-fMRI, processed by GraphSAGE-based graph blocks with GraphNorm and TopK pooling, then passed through an RNN with visit-gap augmentation $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$6. Using focal loss with $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$7 and $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$8, the model reports overall accuracy $$a_{ij} = \exp\left\{-\big\|\mathbf{R}_l^\top (\mathbf{f}^i_l - \mathbf{f}^j_l)\big\|_2^2 \right\},$$9, balanced accuracy $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$0, AUC-ROC $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$1, CN$\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$2MCI conversion accuracy 68.8%, and MCI$\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$3AD conversion accuracy 67.6% (Moghaddami et al., 7 Apr 2026).

These formulations broaden the field from static cognition decoding to multimodal fusion, cognitive graph generation, and temporal modeling of cognitive trajectories.

6. Limitations, misconceptions, and open directions

Taken together, the literature suggests that “cognitive” does not denote a single standardized layer or training recipe. In some papers it refers to latent cognitive-state inference from behavioral graphs; in others it refers to explicit reasoning modules; in connectomics it often denotes graph models whose outputs are cognitive tasks, intelligence scores, or impairment stages. A plausible implication is that the term names a design orientation rather than a uniquely specified architecture.

Several technical limitations recur. Dynamic graph learning can require dense $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$4 pairwise computations; JLGCN explicitly notes dense adjacency construction and memory limits, including subsampling Pubmed to 10,000 nodes (Tang et al., 2019). Exact spectral methods require Laplacian eigendecomposition with $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$5 cost, which SpectralBrainGNN identifies as manageable for $\mathbf{M}_l = \mathbf{R}_l \mathbf{R}_l^\top$6 but not obviously for much larger graphs (Maji et al., 31 Dec 2025). Neuro-symbolic models confront representational and optimization constraints: GRNs inherit adjacency-string scaling and delicate joint training of encoder and differentiable SAT solver (Zopf et al., 2024), while GNNLR restricts explicit reasoning to propositional logic with neuralized NOT and OR operators and identifies first-order logic as future work (Chen et al., 2023).

A second cluster of limitations concerns temporal and causal interpretation. Several connectome-based models operate on static functional connectivity and explicitly point to dynamic fMRI, temporal GNNs, or longitudinal analysis as future directions (Maji et al., 31 Dec 2025, Thapaliya et al., 2023, Mazumder et al., 31 Mar 2026). In educational diagnosis, uncertain edges created by guessing or carelessness motivate explicit edge filtering and reliability-aware graph construction rather than direct acceptance of response logs as ground truth (Shao et al., 2024). In adaptive graph learning, learned relations are symmetric similarity graphs and therefore do not by themselves capture richer causal, temporal, or logical relations (Tang et al., 2019).

The forward directions named in the corpus are correspondingly consistent: sparse or sampling-based graph inference, probabilistic relational reasoning, memory modules, dynamic functional connectivity, multimodal structure–function fusion, more expressive symbolic logic, and larger, more diverse datasets. This suggests that future cognitive graph neural networks will likely be judged not only by predictive accuracy, but also by whether their graph constructions remain interpretable, uncertainty-aware, temporally adequate, and aligned with the cognitive variables they claim to model.