Faculty Sponsor: Raghu Ramanujan, Dan Layman
Machine learning algorithms have been known to exhibit bias in analyzing data and making predictions. Matrix factorization is one such powerful algorithm that is widely used in recommendation systems that, when used without caution, has the potential to perpetuate existing social biases when recommending items to users. This poster investigates whether the phenomenon occurs when matrix factorization is utilized to recommend courses in departments to Davidson College students. Different departments have historically drawn a gender-skewed number of student majors, and this research aims to understand whether a recommendation algorithm could reinforce these tendencies. For this study, recommendations are made based on data about students’ past success, enjoyment and average letter grade in departments in which they have taken a course. Apple’s open-source Turi Create library is used to perform the matrix factorization and to evaluate the results. The matrix factorization algorithm first identifies latent factors between students and departments, reconstructs a user-item matrix to fill in departments not rated by the student, and finally recommends the top-rated departments that the new ratings matrix predicts. The model is not informed about student gender. The poster explores whether a recommendation algorithm could have the potential, in this situation and in others, to influence user behavior in accordance with societal biases and expectations, in a way that may be detrimental to society.