Graph Convolutional Networks: List of Applications You Need To Know

Convolutional Network is a type of Neural Network. Neural Networks are a special kind of deep learning model. Typically, machine learning or deep learning comprises rigorous and expensive algorithms due to its complicated task. Similarly, deep learning models on graphs are even more complicated. Graph Convolutional Networks are primarily used in the purpose of image classification. Learn more about convolutional neural network.

Through the last decade, the application of data science has increased enormously. In this data-rich world, the learning model approach brought great results and accurate predictions. Graphs are useful for many information systems. 

From biological protein interactions to internet connectivity and WorldWideWeb, graphs represent all these systems. Also, implementing neural networks through graphical structure lets the computer understand the properties of an image. This model is one of the most advanced real-world applications of the graph. Let us discuss these algorithms in detail: 

Enrol for the Machine Learning Course from the World’s top Universities. Earn Masters, Executive PGP, or Advanced Certificate Programs to fast-track your career.

How Neural Networks are Built

Neural Networks are one of the most advanced techniques of data science and deep learning. Neural Networks are useful in many applications, from Stock Market prediction to image classification, speech or character recognition and even in sequence analysis.

The first concept of the Neural Network came from biological perspectives. Scientists conducted experiments in which nerves for vision are connected with the hearing centres of the brain. Eventually, the organism learnt to see through the hearing centre of the brain too. Even further experiments proved that every centre of the brain could perform every action.

Approaches started to begin to mimic the human brain informing computer algorithms. So, similarly, computer scientists also thought that there should be a single algorithm which is capable of solving all computer brain learning problems. That is how the neural network came to birth. 

A neural network consists of multiple layers of neurons. Each neuron is typically a graph node. Each neuron of each layer is connected with all the neurons of the next layer through a weighted edge—the weights of the edge act as a coefficient of the layer value calculation.

Through backpropagation, coefficients change to fit the model with sample training examples. Ultimately, a single neuron from the last layer gives the output. In the following image, the structure of a neural network is explained. 

Image Source

Read: Using Convolutional Neural Network For Image Classification

Graph Convolutional Networks 

Convolutional Networks are 3-dimensional neural networks. Most practical uses of Convolutional Neural Networks include image classification and recognition, natural language processing and speech recognition. These models are usually more complex than the usual 2-dimensional neural network models. 

In this architecture, layers of different neurons are assembled. The parameter of dimensions is variable in different layers to make the model recognise parameters. For instance, images are two dimensional, and, in the meantime, the colour of each point also play a crucial role. Hence, three different parameters emerge. To deal with such complexities, Conv Nets play a significant role. 

These many 3d matrices of different dimensions work at multiple levels of the neural network. Eventually, the ‘z’ dimension fit the output parameter of the network. The conveyance of information from one level to another can take place through a variety of different algorithms.

For instance, FC (Fully Connected), Pooling and ReLU are some crucial algorithms regarding this. 

Generally, the node values of the neural network are denoted with, where ‘l’ signifies the layer number. So, a0 is the input matrix. 

a0=X

On the other hand, the last layer node defines the output. Say, there are ‘L’ layers. Therefore aL denotes the output of the neural network.

aL=Y

Image Source

The above image depicts a convolutional neural network in the implementation of image classification. The parameters are set for a dog, cat, bat and bird. 

The node value of a particular internal layer is calculated through previous layer values.

al+1=f(al,θ)

Here, is the adjacency matrix and, f is the defining function. Every graph convolutional network layer can be written using this expression. In this way, a graph convolutional neural network typically works.

Notations and Preliminaries

The basis for comprehending GCN deep learning is graphs and graph signals. Graph signals refer to node-specific attributes, whereas graphs describe things (nodes) and the relationships between them (edges). We can understand the core ideas behind GCNs and their applications by fully comprehending these key ideas.

Graph Fourier Transform

The Graph Fourier Transform provides a mathematical foundation for analyzing graph signals in the frequency domain. This transform allows us to extract useful information from graph-structured data, aiding efficient graph filtering and analysis. It is similar to the Fourier Transform in signal processing.

Graph Filtering

Thanks to graph filtering techniques, we can process graph signals, extract important characteristics, and carry out operations like smoothing, denoising, and improvement. Using graph filtering, GCNs can efficiently extract and represent information from graph-structured data.

Spectral Graph Convolutional Networks

Convolution operations on graph signals are carried out by Spectral Graph Convolutional Neural Networks using the Graph Fourier Transform. These networks may efficiently aggregate data from nearby nodes, enabling potent representation learning on graphs by taking advantage of the eigenvalues and eigenvectors of the graph Laplacian.

Spatial Graph Convolutional Networks

Convolutional networks investigate Spatial Graph Alternative methods for performing graph convolutions. Three broad categories can be used to group them:

Classic CNN-based Spatial GCNs: Architecture of convolutional neural network modification to operate on graphs is one strategy used in spatial GCNs. This entails specifying convolutional operations on graph nodes and edges like CNNs do for standard grid-like architectures. These networks may identify spatial patterns and characteristics in graph data by utilizing local neighbourhood knowledge and shared weights.

Propagation-based Spatial GCNs: The concept of message passing or propagation is the foundation of another strategy in spatial GCNs. This method uses the graph’s connectedness to spread information across nodes iteratively. To update its representation, each node gathers data from its neighbours. This propagation mechanism allows The network to capture spatial dependencies and learn expressive node representations.

Related General Graph Neural Networks: A larger family of Graph Neural Networks (GNNs) that can handle graph-structured data includes spatial GCNs. While Graph Attention Networks (GATs) or Graph Isomorphism Networks (GINs) may emphasize different characteristics, such as attention processes or graph isomorphism invariance, other GNN versions, such as Spatial GCNs, concentrate on spatial information.

Applications of Graph Convolutional Networks

Applications of Graph Convolutional Networks (GCNs) can be categorized into computer vision and natural language processing domains. Here are the key points about their applications:

Computer Vision Applications

Images: Semantic segmentation, object detection, and image categorization are just a few of the image tasks for which GCNs have been effectively applied. By describing images as graphs, where pixels or image regions are nodes and their spatial relationships are edges, GCNs can capture spatial dependencies and contextual information. This enhances the comprehension and processing of images.

GCNs can be utilised for a variety of tasks, such as action recognition, video summaries, and video segmentation. By representing films as spatiotemporal networks, where frames or video segments are nodes and their temporal connections are edges, GCNs may effectively characterise motion dynamics and capture long-range interdependence. This enhances the understanding and interpretation of video.

GCNs are helpful for point clouds, which are representations of 3D data acquired from sensors like LiDAR. With points acting as nodes and edges denoting their spatial interconnections, point clouds can be processed by GCNs as graphs. Point cloud segmentation, object detection, and categorization problems enable applications like autonomous driving, robotics, and 3D scene interpretation.

GCNs can manage mesh data, which represents complex 3D surfaces with vertices, edges, and faces. Using the connection data present in meshes, GCNs may finish meshes and carry out operations such as shape classification, shape generation, and mesh completion. These skills have been applied to computer-aided design, virtual reality, and computer graphics.

• Graph Convolutional Networks generate predictions over physical systems, such as graphs, their interactive approach and applications. GCN also provides accurate information about the properties of real-world entities and physical systems (dynamics of the collision, objects trajectories).  
• GCNs are used to perform image differentiation problems. The model it follows is known as ‘Zero-Shot Learning’. The main motive of this model is to identify an unknown labelled image and group it into known ones. They also gather semantic information of these labels and categorise them.
• GCNs can take a certain length of molecular fingerprints as input and generate predicted molecular structures. MolGAN is one kind of Graph Convolutional network which helps to create new molecular structures with various features in it. In this way, it allows scientists to invent modern molecular structures day by day.
• GCN is applicable for solving various problems related to research operations and combinatorial optimisation applications. Graph Convolutional Networks play a pivotal role in solving salesman problems, quadratic assignment problems, and many more.  With the help of the input graph, it can outclass traditional complex algorithms.

Graph Convolutional Networks for Text Classification

Graph Convolutional Networks are effective for text classification. They present text data as a graph and capture local and global dependencies. While the nodes represent textual elements, the edges represent relationships. These networks employ message passing to collect information from neighboring nodes and capture semantic relationships. They are better than traditional methods because they capture context and achieve accurate classification in sentiment analysis, document categorization, etc.

Karate Club of Zachary

Another significant application of Graph Convolutional Networks is to solve community prediction problems, such as Karate Club of Zachary. This problem is based on the dispute between the administrator and the instructor of the club.

We have to figure out which side every member of the karate club would select. This problem gets resolved by using semi-supervised learning techniques. By using just two labelled nodes, Tobias Jepsen was able to fix the problem and reach near-perfect accuracy in terms of predicting those two communities.

Also Read: Neural Network Project Ideas

Now let’s take a look at the following images and, you would be able to get some insights about the Karate club problem and its proper calculations using Graph Convolutional Networks.

Image Source

Wrapping Up

By reading this article, you would be able to understand what Graph Convolutional Networks are, how Neural Networks are built, a brief idea of GCN and how it works, and various crucial aspects and applications of GCN including Zachary Karate Club problem.

If you want to know more about GCN and its features and benefits, do register at upGrad Education Pvt. Ltd. And IIITB’s Post Graduate and Diploma course on Machine Learning and Artificial Intelligence. This course on Machine Learning and AI is designed for students and working professionals.

The course provides a collection of case studies & assignments, industry mentorship sessions, IIIT Bangalore Alumni status, job placement assistance with top companies, and most importantly, a rich learning experience.