The Context of Machine Learning in Prediction Models (Invited Presentation)

Machine Learning (ML) tools and model structures are powerful methods of solving a range of numerical and data-driven problems, including those which are complex and non-linear. One of the most valuable insights to approaching a project or problem with ML solutions is to understand the context of the problem and application. Context in this sense refers to the beginning problem/question, input dataset(s), and labels (often obtained from the desired output of the model) all the way through to the application of the model itself. In this process, an understanding and statement of biases, statistical relationships, and assumptions in input data should be made along the way and are important to the application of the designed ML model. Sometimes through these steps, it is discovered that more simple approaches work as well as complex ones and that the simpler modeling tools can be to solve the problem. Lastly, understanding which methods are best for analyzing the performance of ML models is not often numerically straight forward but important.

In this presentation, these points of importance will be emphasized in a description of a project that sought to model hourly wildfire radiative power (FRP) into the future given inputs of satellite and numerical weather model data.