Three groups of applications:

  1. Structural scenarios – where data has explicit relational structure, for example, knowledge graph

  2. Non-structural scenarios – data such as images or text where there aren’t any explicit relational structure

  3. Other application scenarios – generative and combinatorial optimisation problems

Structural Scenarios

GNNs are commonly used in social network, recommendation systems, and graph representation. For example, with 1 – Knowledge Graphs , GNNs can be used to tackle the out-of-knowledge-base (OOKB) entity problem in knowledge base completion.

Non-structural Scenarios

Both images and text data doesn’t have explicit relational structure. For text, GNNs can be apply to both sentence-level and word-level text problems. Example text applications include:

  • Text classification – both classic GCN and GAT model. Most methods treat each document or sentence as a graph of words

  • Sequence labelling

  • Neural machine translation – incorporate syntactic or semantic information into the NMT task

  • Relation extraction

  • Event extraction

Other Application Scenarios

There are other interesting applications of GNNs such as generative models and combinatorial optimisation. NetGAN is one of the first neural graph generative model! Many interesting combinatorial problems over graphs are NP-hard problems and research has been applying GNNs model to solve these problems!

Future Work

Shallow Structure

GNNs are always shallow (no more than three layers) in comparison to traditional deep neural networks. Overstacking GNNs will lead to over-smoothing results and although some research has managed to tackle this problem, this still remains as one of the main challenges of GNNs. Future research in being able to effectively stack layers and layers in GNNs will be a strong move.

Dynamic Graphs

Static graphs are stable and more easy to model, however, in real-world, things change. New things are created and old things are destroyed or modified, all of which results in the change of the graphs. GNNs fail to adapt to this dynamic nature of graphs. Dynamic GNN is an active research area and the progression will contribute towards the stability and adaptability of general GNNs.

Non-Structural Scenarios

There are currently no optimal methods in generating graphs from raw unstructured data. Therefore, finding the best graph generation approach will strongly contribute towards the GNN field.


The scalability of GNNs is an issue because of the cost of core computational steps increases as data increases!



Data Scientist

Leave a Reply