BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Memento EPFL// BEGIN:VEVENT SUMMARY:Human Decisions and Machine Predictions DTSTART:20170208T140000 DTEND:20170208T153000 DTSTAMP:20260504T021224Z UID:6861da819ac400b3cbf1e0317c3f5de3341313c602df2bad574f59f4 CATEGORIES:Conferences - Seminars DESCRIPTION:Professor Jure Leskovec\, Stanford University\, Chief Scientis t at Pinterest\, USA\n\nJure Leskovec is As sociate Professor of Computer Science at Stanford University and Chief Sci entist at Pinterest. Computation over massive data is at the heart of his research and has applications in computer science\, social sciences\, econ omics\, marketing\, and healthcare. This research has won several awards i ncluding a Lagrange Prize\, Microsoft Research Faculty Fellowship\, the Al fred P. Sloan Fellowship\, and numerous best paper awards. Leskovec receiv ed his bachelor's degree in computer science from University of Ljubljana\ , Slovenia\, and his PhD in in machine learning from the Carnegie Mellon U niversity and postdoctoral training at Cornell University.\nAbstract\nIn t his talk we examine how machine learning can be used to improve human deci sions---in particular\, on judges deciding whether to jail an arrestee pen ding resolution of their case. This is an ideal application because\, by l aw\, this decision must rely on judges’ prediction of what a defendant w ould do if released. There are several interesting methodological challeng es in this domain. First\, we must solve a selection problem: we do not ob serve what jailed defendants would do if they were released. Second\, judg es’ payoffs may involve more than minimizing crime risk. Algorithmic rec ommendations may reduce crime risk but may not improve the richer total we lfare function. We develop a methodology to deal with these problems and f ind that machine Learning can reduce crime by up to 24.8% with no change i n jailing\, or reduce jail populations by 42.0% with no increase in crime. Such gains can be achieved while simultaneously reducing racial dispariti es as well as reducing all categories of crime\, including the most violen t. We also develop methods to identify reasons for judicial error---judges overfit to unobserved ‘noise’. These findings suggest that machine le arning and prediction tools can be used to understand and improve human de cisions. On the other hand\, they illustrate how this must be a joint acti vity between the design of prediction algorithms and the development of an economic framework that focuses on payoffs\, decisions and selection bias es. LOCATION:IC Faculty\, BC420\, 4th floor STATUS:CONFIRMED END:VEVENT END:VCALENDAR