Images, text, and speech data are usually arranged neatly and are suitable for processing using convolution or sequence models. But in reality, there is still a large amount of "relational" data - social networks, molecular structures, knowledge graphs, transportation networks, paper citation networks - which naturally form a graph: composed of nodes and edges, with irregular structures and varying numbers of neighbors for each node. Graph Neural Network (GNN) is designed for this type of data.
The core operation of GNN is called "message passing": each node first sends a message to its neighbors, and then aggregates the messages it receives to update its own representation. This process is repeated several times, and nodes can gradually "see" a wider range of neighbors, thereby integrating local structural information into their vector representation. In other words, GNN enables each node to learn how to describe itself using 'who am I connected to, who are my neighbors'.
Compared to convolutional networks that can only handle regular grids, GNN has three distinct characteristics: firstly, it can directly handle non Euclidean graph structures; Secondly, it has permutation invariance, and the order of node numbering does not affect the results; Thirdly, the number of parameters is independent of the number of nodes, and the same set of rules can be reused on graphs of any size.
The landing scenarios are very broad: in molecular and drug discovery, atoms are nodes and chemical bonds are edges, and GNN can predict molecular properties and assist in screening candidate drugs; In recommendation systems, users and products form a bipartite graph, and GNN can capture higher-order relationships of "looking and looking"; In financial risk control, it helps identify gang fraud; It can also be seen in traffic prediction, knowledge Q&A, and code analysis.
Limitations also exist: when the image is stacked too deep, it is prone to "oversmoothing", and long-distance information is actually diluted; The training and sampling of large-scale graphs also test engineering skills. In recent years, directions such as Graph Transformer and graph based models have attempted to find a better balance between expressive power and efficiency.
In summary, if CNN is good at "looking at pictures" and RNN is good at "looking at sequences", then GNN is good at "looking at relationships". In today's world where more and more data is organized into networks, it is an essential foundational capability that cannot be bypassed.
【 Reference Source 】 Comprehensive compilation of publicly published literature and reviews on graph neural networks.