Graph Neural Network Papers Define Message Passing, Induction, Attention, and Expressivity Limits
Graph neural networks are often introduced as deep learning on graph data. The paper literature supports a more structured taxonomy: graph learning moved through recurrent graph neural models, spectral convolution, localized graph convolution, inductive neighborhood aggregation, attention-based weighting, and expressivity analysis. This paper synthesizes eight core papers to ask what a GNN claim should specify. The contribution is a message-passing accountability model that separates representation update rule, neighborhood sampling, inductive setting, edge/message semantics, attention weighting, and expressivity limit. The synthesis finds that "uses a GNN" is too coarse for method comparison: GraphSAGE, GCN, GAT, message-passing neural networks, and GIN answer different design questions. The result is a compact literature map for reading GNN papers without collapsing architecture, task setting, and theoretical capacity.
Introduction
Graphs encode entities and relationships, so a graph neural network claim is always about both representation and topology. Early graph neural models established iterative state updates on graphs [[cite:scarselli2009]]. Later spectral and localized graph convolution made graph learning fit modern deep-learning workflows [[cite:bruna2014,kipf2017]].
This paper asks what should be specified when a paper says it uses a GNN. The answer is not a single architecture name. A defensible claim names the message function, aggregation rule, sampling or transductive setting, edge semantics, and known expressivity boundary.
Method
The synthesis selected core method papers that are frequently used to define the modern GNN family. Sources were coded by their primary contribution to the design space: recurrent update, spectral filter, localized convolution, inductive aggregation, attention, message-passing generalization, or expressivity analysis.
Results
The first result is historical layering. Scarselli et al. provide an early graph neural model [[cite:scarselli2009]], while Bruna et al. and Kipf and Welling make graph convolution central to deep-learning practice [[cite:bruna2014,kipf2017]]. These papers are not interchangeable: they answer different representation questions.
The second result is that task setting changes the method claim. GraphSAGE targets inductive representation learning on large graphs by sampling and aggregating neighborhoods [[cite:hamilton2017]]. That is a different accountability claim from a transductive benchmark in which the graph is known during training.
The third result is that aggregation design and expressivity must both be named. GAT learns attention over neighbors [[cite:velickovic2018]], message-passing neural networks generalize update rules across chemistry tasks [[cite:gilmer2017]], and GIN links architecture design to Weisfeiler-Lehman-style expressivity limits [[cite:xu2019]].
Discussion
The literature map shows why "we use a GNN" is not enough. A GCN, GraphSAGE model, GAT, message-passing neural network, and GIN-style architecture place accountability in different parts of the pipeline. For reproducibility, a paper should specify the graph construction, node and edge features, neighborhood rule, aggregation/update functions, training setting, and expressivity caveat.
The limitation is that this synthesis does not benchmark architectures on a common dataset. It reads method papers to identify what must be reported before such a benchmark is meaningful.
Source Boundary and Reporting Checklist
The source boundary is deliberately paper-first: the synthesis uses primary method papers, review papers, and trial or benchmark papers as evidence, and it treats the accountability model as the paper's own inference. For graph-neural-network, the earliest cited source establishes the first durable research claim, while later sources either extend the claim, operationalize it, or restrict its interpretation [[cite:scarselli2009,bruna2014]]. That boundary prevents the synthesis from turning a famous result into an all-purpose slogan.
The reporting checklist below is designed for readers who encounter a new graph-neural-network claim in a paper, preprint, grant proposal, product note, or policy brief. It is intentionally stricter than a summary because a summary can say what the field achieved, while a checklist asks what must be present before the claim can travel to a new context. A source can be important and still be insufficient for a downstream claim if the denominator, measurement method, or use boundary is missing.
The checklist also clarifies the novelty boundary of this article. The cited sources provide the factual claims; this article contributes a reusable reading model that classifies those claims into accountable layers. For example, the model does not assert that every later graph-neural-network paper must cite the same eight references. It asserts that later work should disclose the equivalent evidence layers before asking readers to accept a transferred claim.
A second boundary is temporal. Foundational papers often define the vocabulary of a field, but later papers change the default interpretation by adding scale, new assays, broader databases, harder benchmarks, or negative results. For graph-neural-network, this means the oldest paper in the chain should be read as origin evidence, not as the final statement of operational readiness. Later papers do not erase the origin claim; they add the conditions under which that claim can be reused without overreach.
A third boundary is transfer. A claim can move safely from one setting to another only when the target setting preserves the key assumptions of the cited source. If the setting changes, the new paper has to show why the original mechanism, measurement, or benchmark remains relevant. This is the difference between citation as background and citation as support. Background citations explain why a question matters; support citations carry the actual weight of the claim.
A fourth boundary is failure mode accounting. Every mature literature contains papers that show limits, artifacts, or narrower interpretations. Those papers are not peripheral; they are part of the evidence system because they define what a careful reader should refuse to infer. In this synthesis, the limiting evidence is used to make the central claim more precise, not weaker. A claim that survives stated boundaries is more useful than a broader claim that hides them.
In practice, the checklist should be applied before a claim is used for comparison, funding, deployment, clinical translation, product design, public communication, or policy. The reader should ask whether the new use is repeating the original measurement or merely borrowing its authority. If it is borrowing authority, the new work needs an explicit bridge: same mechanism, same measurement, comparable denominator, and a limitation check. Without that bridge, the citation is informative but not load-bearing.
The article therefore treats graph-neural-network as a case study in disciplined synthesis. It does not attempt to replace specialist reviews, reproduce experiments, or update every downstream paper. Its narrower purpose is to turn a cluster of influential papers into a reusable reading protocol: identify what the papers directly show, identify what later papers changed, and state what must be true before the claim travels beyond its original evidence setting.
This boundary matters because research influence often grows faster than reporting discipline. A method paper can become a benchmark norm; a benchmark norm can become a deployment claim; a deployment claim can become a public narrative. The final cited source in this paper is included partly to keep that chain honest: it either extends the original result into a new setting or shows why the original result needs a narrower interpretation [[cite:li2016]].
Conclusion
Core GNN papers define a design space, not a monolith. Message passing, induction, attention, and expressivity boundaries should travel with every GNN claim. Without those details, model comparisons collapse architecture, data setting, and theoretical capacity into one vague label.