My intern project was to tune the hyperparameters of a probabilistic data structure, the Vector of Counts, that the team proposed to use for forecasting. When The Trade Desk does forecasting, there is a policy that only up to 90 days of historical user data can be stored. This makes long-term forecasting difficult. The Vector of Counts aggregates the data so that more than 90 days of data can be stored while still protecting user privacy. My project helped enable long-term forecasting to understand seasonality and long-term trends.I used Databricks to write code for hyperparameter tuning using Spark Scala, and it is in a GitLab repository for future use. By analyzing the data that was generated when I ran the code to tune the hyperparameters for key Vector of Counts operations, I determined the best hyperparameter values. I wrote several Confluence pages (webpages used by The Trade Desk) documenting the results. The tech lead on my project will continue to work on the project.

I worked in a team that investigated data oversampling techniques to improve the performance of machine learning models that predict the health risk of people with asthma. The patient data consisted of imbalanced binary classification datasets. We surveyed machine learning literature and implemented state-of-the-art oversampling algorithms, and we also implemented our own techniques. I implemented oversampling algorithms in Python, including existing techniques such as Gaussian-based SMOTE and one of our own approaches, Average Neighbor Vector Oversampling. I contributed to a research paper comparing our algorithms to the state-of-the-art oversampling techniques. Our work was accepted at the 2023 National Conference on Undergraduate Research in Wisconsin, where my team and I presented our project.