GNN: Graph Neural Networks

Graph Neural Networks (GNNs)

Strong suggest to read: A Gentle Introduction to Graph Neural Networks; Adam Pearce, et al.; distill; 2021

A Graph Neural Network (GNN) is a type of deep learning model designed to operate directly on graph-structured data by iteratively “passing messages” along edges: each node maintains a feature vector, exchanges information with its neighbors through learnable aggregation functions, and updates its embedding to capture both local connectivity and node attributes; by stacking multiple such layers, a GNN can learn representations for individual nodes (e.g., for node classification), entire graphs (e.g., for graph classification), or edges (e.g., for link prediction), making it highly effective for any task where relationships between entities matter.

!!! Pure Personal Understanding for Quick Understanding the GNN
For CNN, after multiple rounds of convolution by using different kernel like Alex net, the 2D image is encoded into a single vector. So as the GNN. For graphic data, we sort the layers of each nodes based on the graphic sctrcutre. Then we convlution the nodes by using the same kernel which included the nodes from the next layer. The nodes would updated as the sum of nodeweight(nodes adjacent). If there are attention, the attention are further included. Finally, then entire graphic data would be encoded into a single vector as input according to the features. The most straight forward way is take the sum/mean value of all encoded nodes as the final graph representation. During the training, the weight used in the nodes update are update. The attention are also a learnable parameter. Because the vector then input into the neural network, the next steps work exactly the same as CNN.