Testing the Robustness of Machine Learning Models Through Mutations

The reliable performance of machine learning algorithms stands as a critical and foundational concern. Usually, scientific attention centres solely on this aspect when selecting among models. However, in real-world scenarios, datasets are vulnerable to human errors during data input. Consequently, algorithms must display consistency and resilience against such errors. We assert that, especially in real-world applications, the resilience is, akin to the performance, a critical characteristic when selecting one algorithm over another. To address this concern, we propose a novel methodology for assessing model robustness by evaluating models both before and after applying mutations to the dataset. To validate the effectiveness of this methodology, we analyse five commonly used machine learning algorithms in a case study concerning traffic flow forecasting in Madrid. In assessing the robustness of the models, we introduce two metrics derived from well-known regression measurements. The results clearly reveal that the random forest model shows the highest robustness, according to our analysis, and that different models can exhibit very different behaviours in terms of this aspect.