Supervised Machine Learning Techniques to Calculate the Robustness of Networks
Jordi Paillissé, Sergi Bergillos, Martí Madrenys, Eusebi Calle · 2023
The robustness of a transport network (telecommunications, water, electricity, etc.) gives an indication of its resistance to various types of attack (targeted, random, cascading, etc.). This metric is usually calculated analytically, with the help of network simulators. However, this calculation is computationally expensive. In this paper, we propose accelerating this computation leveraging various Machine Learning techniques. We study five different techniques: linear regression, k-nearest neighbors, decision trees, random forest and a multilayer perceptron. We have built and evaluated the accuracy of these five models, using datasets of synthetic telecommunication networks from Topology Zoo, under different percentage of attacked nodes and attack algorithms. Our results show that we can predict the robustness of given a set of input graph metrics with a mean absolute percentage error of 2.36 % for a random forest regressor. In addition, we analyze which metrics are more relevant when performing the prediction, and find that the percentage of attacked nodes is the most important.