[Deep Learning] Why Graph Neural Network?

Introduction

Why do we use graph neural network? In the protein interaction prediction task, the graph neural network (GNN) is often used.

The protein and its interaction are represented as a graph structure. The above is mentioned as a reason for using GNN in ordinary researches. However, we wonder why GNN works better than other machine learning (ML) or deep learning (DL) models, such as random forest or convolutional neural network (CNN).

The goal of this post is to explain why we use GNN by comparing the other model.

Receptive Field

In this section, the background information of the receptive field is described.

Fig1. From left to right, information aggregation at the first layer, second layer, and third layer.

In Figure 1, information aggregation and propagation with the three-layered neural network are shown. The deeper the layer, the more information we can see (or get aggregated information) in a wider area based on the center we are focusing on.

Convolutional Neural Network

Fig 2. One dog and three flowers.

In the CNN case, one of the goals is object recognition. For object recognition, we need to aggregate the spatial information within the picture such as red boxes in the Figure 2.

For recognizing the dog’s face, the background information (landscape) and the flowers are not important things. Thus, the minimum requirement to construct a CNN is to have a receptive field that is larger or equal to the size of the dog’s face.

Graph Neural Network

Fig 3. Protein interaction prediction. The red circle marks the point where we would like to see the interaction.

Figure 3 shows a complex in which some of the two proteins are bound. The red line represents the information transmission relationship between nodes, and the Euclidean distance between nodes can be very long.

Fig 4. The extreme case of information aggregation on graph data

If the Euclidean distance between nodes is very long, but they are closely related to each other, so when the two pieces of information need to be combined, the receptive field should be set very large. Figure 4 shows an extreme case as mentioned above.

However, in the case of using GNN, it is possible to aggregate the information of the two extreme far nodes by skipping the intermediate not important link because the operation is performed by node and edge.