Michael Smith

As a project for a course in my Data Science program at SMU, I, along with two other students, performed a cluster-based random forest classification regression analysis on CitiBike's NYC open data set. This data set consists of millions of rental records of CitiBike bicycles in Manhattan. We settled on the question of, "Can we identify CitiBike subscribers based on rental usage?" Using spectral clustering, we identified key attributes that can be used to identify different types of customers, some of which were subscribers. We were ultimately able to identify subscribers to CitiBike's services with ~73% accuracy. While not stellar, we believe some of the discrepancy lied in false positives, which we identified as non-subscribers that matched travel patterns of subscribers. In which case, we developed a simple model to apply to real-world, real-time data to identify key marketing areas to promote CitiBike's subscription service where it would have the most impact, based on rental locations with high false positive renters.

In this paper, we present a model for employee attrition prediction and discuss the ethical impacts of using such a model within private and public sector organizations. As it is in Human Resource personnel’s best interest to improve retention, implementing statistical and machine learning techniques is the most viable means to attrition abatement. To this end, we examine Office of Personnel Management public sector, Bureau of Labor Statistics public sector, and IBM anonymized private sector employee separation data. Three classification models (Methodologies include Logistic Regression, Random Forest, and K Nearest Neighbor) are trained and tested on these data before selecting our best fit model for attrition prediction. We finally use metrics such as Gini and Permutation Importance to identify the most impactful variables in determining prediction outcome before presenting the ethical ramifications of using such outputs in HR planning.

In the name of scientific progress, an amazing number of data sets have been released as a way to promote collaboration and development of the data science community. This paper critically reviews an array of anonymization theories and techniques and applying them to a fake data set, with the goal of helping others weigh the cost, benefit, and, most importantly, the ethics of anonymizing data. We test these techniques on a fake dataset, taking an in-depth look at with what ease someone can be armed with a small amount of data and yet cause serious ramifications.