This session was given by one of my favorite speakers, Larry Rosenberger. Larry is the head of Analytic R&D for Fair Isaac and the former CEO. Larry is always fun and often good at challenging the audience to think differently. The prompt for this talks was his recent reading of "Blink" by Malcolm Gladwell and his work thinking about how the kind of almost instant decision-making Gladwell described can be combined with the kind of data/results-driven analytics Larry has worked on for years.
Larry's talk revolved around the challenge to the "better decisions from data" mindset of most analysts represented by the kind of rapid cognition and adaptive unconscious decision-making discussed in Blink. While judgmental decision-making can be fallible (we tend to focus on a single idea too early and seek evidence to support it and have trouble managing randomness or considering lots of variables), the evidence that we are good at some kinds of "snap" judgments in some circumstances is strong. Larry proposed a number of ways to combine the two using Bayesian Statistics as a framework. His examples included:
- Using unsupervised learning algorithms to find new outliers and then having experts review them to see which ones are worth tagging as potentially fraudulent for instance.
- Use analytics to analyze and instrument subjective overrides of automated decisions to see if some types of override can be built into the automated decision.
- Move from a Champion/Challenger approach (where one or two new decision models are compared with the existing one to see which is better) to one where learning strategies are applied to the decision model to maximize the range of potentially new strategies considered.
All interesting stuff and mostly way over my head. What it made me think, though, was that the future of analytics is not in reporting or spreadsheets, visualization or cubes. The future is to find new ways to bring human expertise and data analysis techniques together to make learning mechanisms that start "smart" and get "smarter".
