This problem focuses on bringing diversity into an automated Graduate Admissions system. The current problem will help alleviate the problem of human biases in the STEM admissions system. My current research focuses on using a Multi-Armed Bandit system with different types of arm pulls. The novel algorithm I am creating is general enough to be extended to other admissions problems, job hiring systems, and other cohort selection problems.

The insurance industry has the opportunity to expand and update by taking advantage of connected houses and commercial buildings. As IoT becomes more popular and with more houses installing smart divices, the insurance industry could increase profits and mitigate loss more efficiently. See my blog post where I explain my proof of concept of this idea.

Functional Data Analysis (FDA) is a growing field in machine learning which provides a more holistic approach to analysis of functional data that is assumed to have a smooth underlying function. The underlying function can be represented using a basis expansion. We used both the Fourier and the wavelet series expansion to represent our functional data. Traditional machine learning approaches default to treating data as feature vectors. In doing this, they do not leverage the fact that there is a smooth underlying function. For example, derivatives can be taken from the underlying function and more information about the data is gained. In this paper, two different methods of functional data classification are developed and discussed. The first method is classification with principal component kernel regression. The second method is classification with discriminative interpolation. In order to do functional data analysis, traditional machine learning approaches need to be extended to a functional context. For instance, principal component analysis and regression have been extended to functional principal component analysis and functional regression. Finally, we compare these two methods against a kernel support vector machine that treats all data as feature vectors. The classification results are comparable with traditional machine learning approaches.
You can find the poster for this project here and the technical report here.