Machine learning and deep learning methods have made enormous strides in the recent past. GNN is a relatively newer deep learning method that falls into the category of neural networks that work on processing data in diagrams. These algorithms can search for information from the graphs and use the information gathered to predict results.
A graph usually represents data with two components – nodes and edges (which form a connection between two nodes). GNN can be applied to graphs to make predictions at the node level, graph level and edge level.
Better than CNNs?
According to research, CNNs can only work with regular Euclidean data such as images (2D grids) and texts (1D sequences), while these data structures can be viewed as instances of graphs. Graphs, on the other hand, are non-Euclidean and can be used to study and analyze 3D data. GNNs can offer the thought process of the human brain, which distinguishes them from other neural networks.
In standard neural networks, the dependency information is only viewed as a characteristic of nodes. GNNs can propagate the graph structure instead of using it as part of the features.
Research also says that in order to fully represent a graph, all possible orders must be traversed as input to the model such as CNNs and RNNs. However, GNNs can propagate to any node at a time, while the input order of the nodes is ignored.
Applications of GNNs
- Natural Language Processing – Graphics can be an important part of NLP applications. They can be used for text classification, information extraction and answering questions.
- Computer Vision – Although the use of GNN in computer vision is still increasing; much progress has been made. GNN algorithms can be used, among other things, for image classification, human-object interaction and image classification with just a few images.
- Physics – This paper states that “a physical system can be modeled as the objects in the system and pairwise interactions between objects”. GNNs can be used by modeling the objects as nodes and pairwise interactions as edges.
- Chemistry and biology GNNs have applications in molecular fingerprinting, predicting chemical reactions, predicting protein interfaces, drug discovery, biomedical engineering, and more.
- Transportation Networks – GNNs are also used to predict traffic movements, traffic volumes, or road density. The nodes can be the sensors installed on roads, while the distance between pairs of nodes measures the edges.
- Recommendation Systems – Graphics are increasingly used in user interactions with company products.
The paper with the name “Graphic Neural Networks: An Overview of Methods and Applications “ talks about some problems that remain with GNN, although much progress has been made. They are:
- Robustness – As part of the neural network family, GNNs are also vulnerable to enemy attacks. While an attack on images and text focuses on features, adversarial attacks on graphics take additional information into account.
- Interpretability – The paper says it is important to apply GNN models with trustworthy explanations to real world applications. As with Computer Vision and NLP, it is important to consider the interpretability of graphics.
- Graph Pre Training – Neural network models require a large amount of labeled data. It is expensive to have such a large amount of human-tagged data. Therefore, self-supervised models are suggested to guide models to learn from unlabeled data available on web sites. This method has proven itself in CV and NLP. The focus was also on pre-training diagrams, but they come up with different problems and focus on different aspects. As this is an emerging area, it still has issues related to the design of the pre-training assignments and the effectiveness of existing GNN models in learning structure or feature information.
- Complex Graph Structures – Graph structures can be flexible and complex in real-world applications. Various papers have been proposed to deal with complex graph structures such as dynamic or heterogeneous graphs.
GNNs can be a very important area in machine learning in the years to come. If the problems listed above can be addressed, then they can be used on big problems as well.
Subscribe to our newsletter
Get the latest updates and relevant offers by sharing your email.