主题：Algorithms for Predictive Analytics: Communication, Privacy and Weights
摘要：
Substantial discussion within policy, legal and industry circles remains on how to regulate machine learning algorithms that are increasingly playing an important role in the lives of individuals. In this paper, we first show that when generating a prediction, both econometric and machine learning strategies respectively place a different weight on each observation of the outcome variable in the underlying training dataset. Our theoretical investigation shows that regression based strategies including those that involve penalization or model uncertainty generate unbounded weights, whereas numerous tree based strategies generate weights that are bounded within the unit interval, ensuring any prediction is over the support of the response variable. In contrast, boosting and (least squares) support vector regression allow for weights outside of the unit interval, thereby allowing a forecast to potentially better extrapolate outside of the support of the response variable. Using data from Lehrer and Xie (2022), we illustrate how to calculate these weights for making predictions both within and outside the support of the training data. We additionally examine whether these weights are associated with variable importance metrics, thereby providing an understanding of how algorithms extract information. Our investigation highlights how prediction weights can help determine the credibility of a prediction or forecast with cross-sectional data. Last, from a policy perspective we argue that an examination of these weights appears consistent with companies complying with the European Union’s General Data Protection Regulation and does not require reporting the technical details underlying the functioning of a complex machine learning algorithm
主讲人：Steven Lehrer
主讲人介绍：Steven Lehrer is a Professor of Economics at Queen’s University. Professor Lehrer also serves as an associate editor for the Journal of the Royal Statistical Society, Canadian Public Policy and Empirical Economics. His work has appeared in the Management Science, Review of Economics and Statistics, Review of Economic Studies, American Political Science Review among other outlets. His research evaluating universal childcare policies was awarded the 2013 John Vandekamp prize for best article in Canadian Public Policy and his work on genetic lotteries received the 2009 Victor R. Fuchs Research Award for his best paper with the potential to spawn new research in an underdeveloped area of health economics or health policy. His research interests include health economics, economics of education and experimental economics, causal inference and genetic information in social science analysis.
时间：2022.6.28 上午10：00
地点：腾讯会议 733 139 2055