Definition

Graph neural networks (GNNs) are a class of deep learning methods designed tow work directly with data structured as graphs.

Graphs are a powerful way to represent complex relationships between entities

Traditional neural networks like CNNs excel at processing grid-like data. GNNs unlock analysis of data where relationship between data points is as important as the data itself.

Key Applications

• NLP
• Drug discovery
• Recommender system
• Traffic prediction
• Computer vision

Importance

• Relational Data Power - Many real-world problems have rich connections that GNNs can naturally model, unlike traditional machine learning methods.
• Permutation Invariance - GNNs are not sensitive to the order in which the graph data is presented, making them very flexible.
• Inductive Learning - GNNs can generalize to unseen graphs, a key advantage for problems where data structures change over time.

Function

• Node Embeddings - Each node (entity) in the graph is assigned an initial vector representation called an embedding. This captures its features.
• Message Passing - Nodes repeatedly exchange messages with their neighbors. These messages update a node's understanding of its place in the graph.
• Aggregation - A node aggregates messages from its neighbors, combining this information with its own embedding. This forms a new, refined embedding.
• Iterative Refinement - Message passing and aggregation happen multiple times. With each teration, a node's representation becomes more informed by its surroundings in the graph.
• Tasks - This process leads to highly expressive representations of nodes, edges, or the whole graph, used for tasks.