Graph Representation Learning is the task of effectively summarizing the structure of a graph in a low dimensional embedding. With the rise of deep learning, researchers have come up with various architectures that involve the use of neural networks for graph representation learning. We call such architectures Graph Neural Networks.
Before getting into the specifics of GNNs, we first understand the motivation behind doing so. One might ask that can we not use our current deep learning architectures like CNN and RNN for graph representational learning. However, it can be observed that CNNs specialize in grid-like structure whereas RNNs specialize in sequential data. In contrast, graphs are more unstructured and would benefit from special methods that can learn in such a setting.
Graphs are everywhere!
Various data stores have an inherent graphical structure and leveraging this data can have a lot of impact. Following are various domains where GNNs could have an impact:
These networks can be leveraged for better recommendation systems, understanding the economic impact of various components, drug discovery, understanding how neurons in the brain function, and many other applications.