Home / Curated / Architectures / Gated Graph Neural Network

Gated Graph Neural Network

Architectures AIMech:GatedGraphNeuralNetwork DRAFT

Aliases: GG-NN, Gated Graph Neural Networks

A graph neural architecture that initializes node states from annotations, repeatedly combines edge-type- and direction-specific neighbor messages with shared gated recurrent updates for a fixed propagation depth, and reads out node or graph predictions.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 599 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 4, 1, 1 lines. All distinct match contents were inspected. Semantic matching covered 49 AIO and 106 native definitions, with every matched definition read. 6 full annotated AIO node/edge/generated contexts and 4 full native records were inspected. The named Section3 GG-NN combines typed directed graph propagation with fixed-depth gated recurrence and node/graph readouts. It is distinct from the Section4 GGS-NN extension, contraction-based original GNN, generic GRU units, CommNet, Interaction Network and GAT. An earlier discovery lead exists but no curated identity does; recover its addressed disposition explicitly.

Claims & Grounded Evidence

6 Claims

Mechanism #typed-directed-gated-propagation

Initialize each node hidden vector by placing its supplied annotation in a prefix and padding the remaining coordinates with zero. At each of a fixed number of propagation steps, sum transformed neighboring hidden states using learned matrices shared by edge type and direction, concatenate incoming and outgoing aggregates, and update each node through a shared gated recurrent unit. Opposite directions may use different matrices; the graph and edge types are supplied inputs.

supports Source: Gated Graph Sequence Neural Networks · Locator: Section 3 and Equations 1–6, PDF pp. 3–4; Figure 1, p. 4

Scope #gate-and-reset-conventions

The displayed update is h_new=(1-z)*h_previous+z*h_candidate, so z weights the candidate in this paper. The candidate uses U*(r*h_previous), applying the reset vector before the recurrent matrix. Applying reset after projection, (r*(U*h_previous)), is generally different for a nondiagonal matrix. These source-specific conventions are not a claim that every software GRU API has identical parameterization or semantics.

supports Source: Gated Graph Sequence Neural Networks · Locator: Section 3.2, Equations 3–6, PDF p. 3; direct constructed gate and matrix examples

Mechanism #node-and-graph-readouts

After propagation, a node-selection head scores each node from its final hidden state and original annotation, then normalizes scores across nodes. A separate whole-graph head sums elementwise products of sigmoid gates and transformed node states before an outer tanh; the paper permits replacing that outer tanh by the identity. Its sigmoid gates are not a softmax distribution over nodes. Shared functions and commutative sums give equivariant node representations and invariant graph readouts under consistent relabeling, rather than an unchanged ordered array of node outputs.

supports Source: Gated Graph Sequence Neural Networks · Locator: Section 3.3, Equation 7, PDF p. 4; direct constructed permutation and gate-weight examples

Limitation #finite-propagation-boundaries

A propagation update exchanges information across adjacent graph nodes, with incoming and outgoing channels both available. With local annotations and fixed propagation depth, a node cannot depend through this propagation path on arbitrarily distant annotations; a final global readout can separately aggregate all nodes. Sum pooling can map different message multisets to the same aggregate. The paper does not establish universal graph discrimination, guaranteed long-range gradient preservation or convergence merely from using gates.

supports Source: Gated Graph Sequence Neural Networks · Locator: Section 3.2, Equations 2–6 and Figure 1, PDF pp. 3–4; Section 3.3, p. 4; direct locality and sum consequences

Scope #unrolled-training-not-fixed-point

GG-NN unrolls a fixed number of gated propagation steps and trains by backpropagation through time. This differs from the prior fixed-point GNN training discussed in Section 2, which relies on contraction for convergence. Removing that requirement permits different learned dynamics but does not certify that repeated GG-NN updates converge. The named Gated Graph Sequence Neural Network extension adds output-sequence and annotation-update machinery beyond this base architecture.

supports Source: Gated Graph Sequence Neural Networks · Locator: Sections 2–4, PDF pp. 2–5; Appendix A, pp. 17–18

Limitation #symbolic-task-evidence

The bAbI experiments use the symbolic data format to construct entity/relation graphs and mark question arguments; they do not establish end-to-end natural-language parsing. Separate question-type models and task-specific dimensions are used, with five propagation steps and varying training-set sizes. Table 1 reports results on selected tasks, including four separate models for task 4. Graph conversion loses temporal order and does not represent arbitrary higher-arity facts. Sequence path and program-verification results use the additional GGS-NN machinery and are not attributed to the base GG-NN alone.

supports Source: Gated Graph Sequence Neural Networks · Locator: Section 5.1, PDF pp. 5–8, Tables 1–3; Discussion, pp. 12–13; Appendix B, pp. 18–19

⚠️ Curated Limitations & Negative Results

The base GG-NN is distinct from the same paper’s GGS-NN sequence extension, the prior contraction-based GNN, named GCN, and generic GRU components.
The supplied graph may be static while its hidden-state computation is recurrent; this does not require an observed temporal input sequence.
The inspected version is arXiv1511.05493v4, dated 22September2017, bearing an ICLR2016 header. All20pages, including appendices, were read. Author code and other versions were not inspected.
Tiny chosen calculations inspect algebraic boundaries only. No model training, benchmark reproduction, convergence proof or empirical performance certification was performed. DRAFT remains.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	broad_match	Learned graph message transformations, gated hidden updates and neural readouts form a particular artificial neural architecture; this target is broader.
`AIO:GatedRecurrentUnit` Gated Recurrent Unit	related_match	GRU-style updates are components of the graph architecture, not an equivalent whole-model identity. The mapping does not adopt the pinned target’s LSTM parentage as a scientific assertion or equate framework-specific GRU variants.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`li2016`	Gated Graph Sequence Neural Networks Yujia Li, Daniel Tarlow, Marc Brockschmidt, Richard Zemel	primary paper full_text	ID: arXiv:1511.05493v4 sha256:7053a744926a…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T09:44:22Z

Updated: 2026-10-13T09:53:49Z

Source File: data/curated/GatedGraphNeuralNetwork.yaml