Introduction to GNN

GNN：graph neural network Figure neural network
because GNN Powerful modeling function between graph nodes , It makes a breakthrough in the research field related to graph analysis . Figure neural network （GNN） It is a kind of method based on deep learning to deal with graph information .

1. Application field

Because of its good performance and interpretability , It has been widely used in various fields . It covers the recommendation system 、 Combinatorial optimization 、 Computer vision 、 Physics / Chemistry and drug discovery .
Recommendation system ： The graph structure data comes from the context of the interaction between users and products on the e-commerce platform , therefore , Many companies use GNN Make product recommendations . A standard case is to model the interaction between users and goods , Then learning nodes are embedded with some form of negative sampling loss , And use KNN The index retrieves similar products of a given user in real time .
Combinatorial optimization
Combinatorial optimization ： Combinatorial optimization （combinatorial optimization, CO） The solution of the problem is finance 、 logistics 、 energy 、 The key in life science and hardware design . Most of these problems are represented by graphs . Google brain team uses GNN Optimized new hardware （ Such as Google Of TPU） Power consumption of chip block 、 Area and performance . A computer chip can be understood as a diagram composed of memory and logic components , Each diagram is represented by the coordinates and types of its components . Determine the location of each component , At the same time, comply with the restrictions of density and wiring congestion , It's a laborious process , But it is still the focus of electrical engineers . Google brain team uses GNN Models and strategies and benefits reinforcement learning （RL） A combination of functions , Generate optimized circuit chip layout , Even better than manually designed hardware layout .
Computer vision ： One way to perceive an image is through a scene graph （ The paper 《Scene Graph Generation by Iterative Message Passing》）, That is, the objects that appear in the image and the set of relationships between them . Scene map has been used in image retrieval 、 Understanding and reasoning 、 Subtitle generation 、 Visual question answering and image generation have been applied .
Physics / chemical ： Life science benefits from representing the interaction between particles or molecules as a graph , And then use GNN Predict the properties of such systems .
Drug discovery ： In biology , Graphs can be represented as interactions of different scales . At the molecular level , The edges of the graph can be the bonds between atoms in molecules or the interactions between amino acid residues in proteins . And on a larger scale , Graphs can represent more complex structures （ Like protein 、mRNA Or metabolites ） The interaction between . According to a specific level of abstraction , These graphs can be used for target recognition 、 Prediction of molecular properties 、 High throughput screening 、 New drug design 、 Protein engineering and drug reuse .
Knowledge map ： The knowledge map itself is also a graph model .
Road traffic ： Dynamic flow prediction

2. chart

The basic structure ： Node and the edge that represents the connection between nodes .
The purpose of using graph neural network is to integrate features .
Vertex（or Node）embedding Node embedding , Expressed as a vector .
The adjacency matrix of a graph ： Indicates the adjacency relationship between nodes