Developing regulations and informing individuals on how machine learning algorithms make decisions requires a transparent metric that provides interpretable information. We propose prediction weights that measure the contribution of each observation’s outcome variable in the underlying training dataset when generating a prediction. We derive prediction weight formulas for popular econometric methods and machine learning strategies. Using data from the film industry we demonstrate how prediction weights can be communicated without providing the underlying technical details of complex machine learning algorithms. Prediction weights provide an interpretable understanding of how algorithms extract information from data and can assist in determining their credibility.
DATE: Friday, November 10, 2023
TIME: 3:30-5:00 p.m.
LOCATION: Fronczak 444