When to Stop-and-Frisk

This project was intended to demonstrate the bias that machine learning algorithms may learn. We wanted to explore the impact bias has on these algorithms and explore the methods that help correct these biases. Given data about stop-and-frisks instances, we ran several different models on the data and maximized the accuracies achieved by these models. After tuning the hyper parameters of the models, we decided to explore logistic regression, which allows us to manipulate the algorithm a bit more than the other models. One means of achieving algorithmic fairness is the idea of statistical parity, where the percentage of drivers of a certain demographic in our dataset is equal to the percentage of individuals of that demographic in the overall population. Overall, we were able to achieve statistical parity while maintaining a consistent accuracy. Thus we learned that we should be aware of the inherent bias in data regarding people and that algorithms can actually learn to be biased and that we, as members of a larger society, should attempt to correct these biases.