This article analyzes three most common approaches to the GNN architecture implementation on the AWS cloud for the use case of the risk assessment in the insurance area. The paper is split to several chapters, with the first one being the overview of 3 approaches to the GNN architecture, the second one describing prerequisites for the implementation, and finally development of the approaches on the cloud infrastructure, testing them on graph insurance data and comparison of all the approaches to select the most suitable for the risk assessment task.
The initial chapter introduces the three architectural approaches to GNN implementation being respectively Graph Convolutional Network (GCN), Graph Attention Network (GAT) and GraphSAGE (Graph Sample And AGgregatE). To conclude the chapter, it is decided to proceed with the further implementation of all three models on the AWS infrastructure and analyze the outputs on the same graph data to select the best suit for the risk assessment use case.
Then the article proceeds with considering the specifics of a realization of risk assessment in insurance on top of cloud infrastructure and preparing the data to use it for the GNN training and testing. After the analysis of the use case, it is decided to focus on only on the individuals’ insurance. The main goal is to analyze the unique properties of every human which can affect the risk of insuring them as well as their connections with other individuals.
Further along, the development of all three approaches for risk assessment solution is described with first being GCN, then GAT and finally GraphSage. The models are then trained, tested and the output analysis is performed. Considering the analysis results, GAT and GraphSage provide the most correct results maintaining the test accuracy. However, considering model statistics, it is found that GraphSage has more distinct probabilities and additional insights through feature importance analysis which makes it the best fit for the risk assessment use case.
The article concludes by stating that out of all three analyzed architectures the most suitable for the risk assessment task is the GraphSAGE with a slight difference between this model and GAT, which will be used for further analysis and improvements. Furthermore, the article mentions a few steps for the potential future improvements of the models, which include using class weights or oversampling techniques to ensure the best performance, also mentioning the experiments that can be done with deeper architectures or different GNN layers. The last but not the least would be to focus on the testing and training on the larger dataset to make it more applicable for real-world applications.
- Danish, A., Xie, Y., Umair, S. B. (2021). Insurers’ risk management as a business process: a pros- pective competitive advantage or not? Emerald, 31 (3), 345–366. DOI: https://doi.org/10.1108/EJMBE-08-2021-0221
- Sanchez-Lengeling, et al. (2021), A Gentle Introduction to Graph Neural Networks, Distill. DOI: https://doi.org/10.23915/distill.00033
- Uzair, A. B., Tang, H., Guilu, W., Marjan, S., Hussain, A. (2023). Deep Learning with Graph Convolutional Networks: An Overview and Latest Applications in Computational Intelligence. International Journal of Intelligent Systems. DOI: https://doi.org/10.1155/2023/8342104
- Kipf, T. N., Welling, M. (2016) Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 DOI: https://doi.org/10.48550/arXiv.1609.02907.
- 5. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y. (2017) Graph Attention Networks. arXiv preprint arXiv:1710.10903v3. DOI: https://doi.org/10.48550/arXiv.1710.10903.
- Totaro, S., Hussain, A., Scardapane, S. (2020). A non-parametric softmax for improving neural attention in time-series forecasting. Neurocomputing, 381, 177–185. DOI: https://doi.org/10.1016/j.neucom.2019.10.084
- Zeng, H., Zhang, M., Xia, Y., Srivastava, A., Malevich, A., Kannan, R., Prasanna, V., Jin, L., Chen, R. (2022). Decoupling the depth and score of Graph Neural Networks. arXiv preprint arXiv: 2201.07858v1. DOI: http://arxiv.org/abs/2201.07858v1
- Liu, T., Chen, Y., Li, D., Wu, C., Zhu, Y., He, J., Peng, Y., Chen, H., Guo, G. (2021). BGL: GPU- Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing. arXiv preprint arXiv: 2112.08541. DOI: https://doi.org/10.48550/arXiv.2112.08541
- Tian, Y. (2022). The World of Graph Databases from an Industry Perspective. arXiv preprint arXiv:2211.13170. DOI: https://doi.org/10.48550/arXiv.2211.13170
- Wang, Z., Ioannidis, V. (2022). How AWS uses graph neural networks to meet customer needs. Amazon Science. https://www.amazon.science/blog/how-aws-uses-graph-neural-networks-to-meet-customer-needs
- Zhang, Z. (2018). Improved Adam Optimizer for Deep Neural Networks. 2018 IEEE/ACM 26th International Symposium on Quality of Service (IWQoS). DOI: 10.1109/IWQoS.2018.8624183.
- Cordonnier, J., Loukas, A., Jaggi, M. (2021). Multi-Head Attention: Collaborate Instead of Concatenate. arXiv preprint arXiv:2006.16362. DOI: https://doi.org/10.48550/arXiv.2006.16362
- Xu, D., Ruan, C., Korpeoglu, E., Kumar, S., Achan, K. (2020). Inductive Representation Learning on Temporal Graphs. arXiv preprint arXiv:2002.07962. DOI: https://doi.org/10.48550/arXiv.2002.07962