Robertson, Duncan, et al. "Crime risk forecasting.”
U.S. Patent No. 9,129,219. 8 Sep. 2015.
From Jun - Sept 2013, I worked at Palantir Technologies researching models to predict the time and location of future crimes. We used ensembles of NNs, SVMs, and logistic regressions.
This work involved encoding both cyclical features (time of day, day of week, week of year) spatial features (x, y blocks within a city) and categorical features (type of crime), and learning different models on a crime-by-crime basis. Learning needed to be data-efficient because some crime types (e.g. homicides) are extremely sparse but important. These models were developed in partnership with a major police department and determined by them to be high quality.
James Green-Armytage, Nicolaus Tideman, and Rafael Cosman. "Statistical evaluation of voting rules.”
Social Choice and Welfare 46.1 (2016): 183-212.
I spent June-August 2012 researching properties of voting systems with Professor Nicolaus Tideman at Virginia Tech. This involved inventing and implementing (in Python) algorithms to measure the manipulability of diverse voting rules including Plurality, IRV, and Minimax. Our team’s goal was to answer the following question:
Which voting rules best encourage voters to give honest ratings of candidates?
Rafael Cosman, Nicolaus Tideman, and James Green-Armytage. "A Three‐Dimensional Framework for Locating Voting Rules.”
Public Choice Society Conference, Mar. 2014
This work generalizes 7 different well-documented voting rules into a single continuous space. Each point in this space is a unique voting rule and the dimensions describe three tradeoffs between voting rules. Arrow's Impossibility Theorem suggests that there is no perfect voting rule - this generalization can help researchers to search along parts of the Pareto frontier.