PAPERS-07-008 · Paper — Message Passing Neural Networks (MPNN)

The paper

The paper “Message Passing Neural Networks” (Gilmer, Schoenholz, Riley, Vinyals, and Dahl, 2017, NeurIPS) is considered a landmark because it unified many earlier graph neural network variants under a single, clear framework, making the field much easier to compare and build upon.

Before this paper

Before this work, a variety of graph neural networks existed, each described with its own notation and terminology. This diversity made it hard for researchers to see connections between methods. The lack of a common language limited systematic progress and comparison. As a result, the field struggled with fragmented understanding and slowed development.

The key idea

The core insight is that most graph neural networks are doing the same thing: passing messages along edges. By formalizing this as a message‑passing process, the authors provided a single, unified view. This view captures how nodes exchange information and update their states. It shows that the differences between prior models lie mainly in the specific message and update functions they choose.

How it works

An MPNN works in rounds. In each round every node creates a message based on its current feature vector and those of its neighbors. All messages are sent to the neighboring nodes. Each node then aggregates the incoming messages (for example by summation) and uses an update function to produce a new node state. After several rounds, a readout function gathers all node states to produce a graph‑level output, such as a property prediction. Existing GNNs can be recovered by plugging in particular choices for the message, aggregation, and update steps.

See it — message, update, read out

flowchart LR
  A["Nodes send<br/>messages"] --> B["Update each<br/>node state"]
  B --> C["Repeat a few<br/>rounds"]
  C --> D["Readout:<br/>graph-level output"]

What they showed

The authors demonstrated that, when framed as message passing, neural networks achieve strong performance on predicting quantum‑chemical properties of molecules in the QM9 benchmark, setting new strong results for that task.

What it changed

This paper gave the community a shared language for describing graph neural networks, turning message passing into the standard mental model. It helped researchers compare, combine, and extend prior methods more easily. The unified framework quickly became the foundation for most subsequent GNN work. Consequently, the field moved forward with clearer communication and faster innovation. Next: the frontier — graph transformers and geometric deep learning.

The one-line takeaway

Message passing is the unifying principle behind virtually all modern graph neural networks.

Read it / meet it

Reading the original paper reveals the elegant simplicity behind many complex GNN variants. Its influence is visible today in every new graph‑learning model, and you’ll see its legacy whenever practitioners discuss building GNNs with message‑passing steps.