Machine Learning and Computational Statistics

Felipe N. Ducau, Michael Higgins, Sebastián Brarda (New York University)

Project Advisor, Ph.D. Kush Varshney (IBM Research)

Online education has become one of the main education formats globally. It has many advan- tages over traditional education like lower costs, scalability and convenience. However it lacks the personalization of traditional education.

It is very challenging to design a system that adapts to the different student skills and weaknesses. The pace at which different students learn varies considerably across different topics and exercise types. By training a machine to understand how well a particular student has learned a particular topic, we could choose which is the most appropriate question to show next in order to optimize her rate of learning. Thus we strive to predict as accurately as possible the prob- ability that a student answers correctly any given question given his/her past performance. If the probability of answering correctly a certain question is very high, the problem would be too easy - showing that problem to the student would be a waste of time. On the other hand, if the problem had a very low probability of success, it would be too difficult, generating frustration without assisting learning.

In this project, we focused on generating calibrated probability estimates for the mentioned task and proposing a personalization system based on these probabilities. The goal was to create a strong probabilistic model using novel feature creation and leverage this model to create an effective recommendation system.

Results and technical description here