
Kyle G
- Research Program Mentor
PhD Doctor of Philosophy candidate
Expertise
Nonconvex optimization, machine learning, signal processing, online and scalable algorithms
Project ideas
Movie Titles Recommendation Engine
Modern day streaming platforms (e.g. Netflix, Hulu, Disney+, etc.) seek to model user preferences for movies/TV shows in order to better recommend new titles. If we collect a matrix of users to movie titles, with the entries being the user ratings of the viewed titles (e.g. 1-5 "stars"), the matrix is highly incomplete, i.e. most of the entries (90-95%) are missing since most viewers have only watched a handful of titles on the platform. With so much missing data, it might seem hopeless. Nevertheless, we can build algorithms to learn a low-dimensional model that allows us to complete and predict the missing entries. From this model, we can also learn factors that explain user preferences for titles (e.g. does user X prefer rom-coms? Marvel movies?). How accurately can we predict user ratings based on the algorithm we design? Can we cluster users together based on their preferences? Can we build "online algorithms" that update their estimates from the addition of new users, rather than computing the entire model from scratch each time?