Page 14 - TECH MAGAZINE CSE

AN Technical Magazine
VESHAN Maharaja Agrasen Institute of Technology

FACULTY
CORNER

Ms. Karuna Midha

Department of Computer Science &
Engineering

Graph Neural Networks (GNNs): Deep Learning for Graph-Structured
Data

Graph Neural Networks (GNNs) and Deep Learning for Graph Data are essential tools for
analysing and learning from data organized as graphs, including knowledge graphs, social
networks, and chemical structures. GNNs are specifically made to handle irregular and
non-Euclidean data represented by graphs, whereas traditional deep learning approaches
like CNNs and RNNs are made for grid-like data, such as images or sequences.
GNNs are neural networks made to learn from graph-structured data by combining and
converting input from the graph's nodes and edges to create meaningful representations
for the nodes, edges, or graph as a whole.

Key Features:
·Node Features: Every node in the graph may be connected with features, such as an

atom's properties or a social network user's attributes.
·Edge Features: In a similar manner, edges may have characteristics (such as the kind of link
between atoms or the type of relationship between individuals).
·Graph Structure: An essential component of GNNs is the adjacency matrix, which shows
how connected nodes are to one another.
·Message Passing: At each layer of the GNN, nodes pass information to their neighbours.
This message can include the node’s current feature vector and any additional edge
information.

GNN Architecture Overview:
The fundamental concept of GNNs is to aggregate data from its neighbours—that is,
neighbouring nodes connected by edges to repeatedly update each node's representation.
Following multiple aggregation cycles, each node's representation of each node encodes
data from its larger network.