Hazardous asteroids forecast via Markov Random Fields: A case study for Explainable Artificial Intelligence (XAI)

The recent interest in machine learning (ML) techniques has raised concerns about their interpretability, especially if they are to be used in physics. In physics, a good prediction should be justified by a model that connects it to fundamental principles. Graphical methods (i.e. Markov Random Field) are a good compromise between interpretability and accurate forecasting. Our study used the asteroid's hazardousness data provided by CNEOS as a dataset since the underlying physics is well-known. Therefore, we can evaluate how well these methods capture physical laws. In addition, we considered other ML algorithms like Random Forest and Support Vector Machines to have a baseline for accuracy and interpretability. Our findings suggest that graphical methods are a viable choice, as they provide a highly interpretable model with sufficient accuracy.