Graph neural networks (GNN) apply the predictive power of deep learning to data structured as objects and their relationships, represented as points connected by lines in a graphCITE:E1. NVIDIA (輝達) describes the data points in a GNN as nodes, linked by lines called edges, with the graph's elements expressed mathematically so that machine learning algorithms can generate useful predictions at the level of individual nodes, edges, or entire graphsCITE:E2.
GNNs use a mechanism called message passing to embed information about a node's neighbors directly into that nodeCITE:E3. Distill formally describes a GNN as an optimizable transformation applied to all of a graph's attributes — nodes, edges, and global context — that preserves the graph's symmetries, a property known as permutation invarianceCITE:E4.
Researchers Wu and colleagues propose a taxonomy that divides state-of-the-art graph neural networks into four categoriesCITE:E6.
| # | GNN category |
|---|---|
| 1 | Recurrent graph neural networks |
| 2 | Convolutional graph neural networks |
| 3 | Graph autoencoders |
| 4 | Spatial-temporal graph neural networks |
The four-category frameworkCITE:E6 gives the message-passing mechanism and permutation-invariance property described aboveCITE:E3CITE:E4 a structured set of architectural families to operate within.
Distill reports that practical GNN applications are already emerging in areas including antibacterial drug discovery, physics simulations, fake news detection, traffic prediction, and recommendation systemsCITE:E5.
Researchers Gilmer and colleagues state that supervised learning on molecules holds significant potential for use in chemistry, drug discovery, and materials scienceCITE:E7.
Taken together, the cited sources describe a single underlying mechanism — message passing that embeds neighbor information into each node while preserving permutation invarianceCITE:E3CITE:E4 — that has been organized into four architectural familiesCITE:E6 and is already surfacing in applications from antibacterial drug discovery, physics simulation, and fake news detection to traffic prediction and recommendation systemsCITE:E5. The antibacterial drug discovery application named among these use casesCITE:E5 lines up with the specific potential researchers describe for supervised learning on molecules in chemistry, drug discovery, and materials scienceCITE:E7, pointing to molecular-level learning as one of the more concrete near-term applications of the technique documented across these sources.
Graph neural networks (GNN) apply the predictive power of deep learning to data structured as objects and their relationships, represented as points connected b…
GNNs use a mechanism called message passing to embed information about a node's neighbors directly into that nodeCITE:E3.
Researchers Wu and colleagues propose a taxonomy that divides state-of-the-art graph neural networks into four categoriesCITE:E6.
Distill reports that practical GNN applications are already emerging in areas including antibacterial drug discovery, physics simulations, fake news detection, …
The technically load-bearing detail here is permutation invariance: because a GNN's transformation preserves graph symmetries rather than depending on node ordering, the same message-passing mechanism can be applied to a molecule graph as readily as to a traffic network or a recommendation graph, which is exactly the span of use cases named — antibacterial drug discovery, physics simulation, fake news detection, traffic prediction, recommendation systems, and molecular supervised learning for chemistry and materials science. That breadth is organized, not accidental: the four-category taxonomy (recurrent, convolutional, autoencoder, spatial-temporal) gives each application domain a matching architectural family to draw on rather than forcing one design to fit all graph shapes. The metric worth watching next is which of these four categories ends up dominant specifically for molecule-level tasks, since the drug-discovery and materials-science potential described here is stated as promise rather than as a settled architectural choice.
Google announced a €13 billion ($15.1 billion) investment in Finnish AI infrastructure to be deployed in 2027–2028, its largest single investment in Europe. The plan covers new data centers in Kajaani, Muhos, and Vaala plus an expansion in Hamina, backed by a 22-year power deal for 50% of Loviisa nuclear plant output and projected to support 37,000 construction-phase jobs and 7,000 permanent roles.
Chenbro (勤誠, 8210) reported August 2026 revenue of NT$2.457 billion, roughly flat month-on-month (-0.2%) but up 40.5% from a year earlier. Cumulative January–August revenue reached NT$19.848 billion, up 51.7% year-on-year.
Chang Hwa Bank (TWSE: 2801) reported cumulative after-tax profit of NT$15.567 billion for January through August, up 23.21% year-on-year with EPS of NT$1.29, a record for the period. The bank attributed the gain to loan, deposit, and wealth management momentum, and August alone delivered NT$2.916 billion in profit, up 29.16% year-on-year.