SEM217: Ram Akella, University of California, Berkeley School of Information and TIM/CITRIS/UCSC: Dynamic Multi-modal and Real-Time Causal Predictions and Risks

Tuesday, September 20th @ 11:00-12:30 PM (639 Evans Hall)

Dynamic Multi-modal and Real-Time Causal Predictions and Risks

Ram Akella, University of California, Berkeley School of Information and TIM/CITRIS/UCSC

There are three major trends in prediction and risk analytics. We describe our research on two fronts and speculate on the third. We do this in the context of healthcare analytics and computational advertising at Silicon Valley firms. We first describe prediction and risk analytics using combined multi-modal numerical data (from vitals and labs) and text data (from notations by doctors and nurses). We describe and analyze dynamic models of patient mortality probabilities, and integrate novel topic modeling to account for topic constraints, and demonstrate superior performance on Intensive Care Unit (ICU) data. We then describe novel experimental design and estimation methods for advertising, using a targeting engine, in the presence of auctions, and the significant benefits in performance versus using (invalid) standard A/B testing (in Silicon Valley firms). We finally speculate on integrated machine learning and causal statistics, and the role of deep learning and reinforcement learning, in these and other problems, including sensor-based IoT in water analytics.