With the advent of massive online open courses, platforms have collected millions of logs from students over questions. Educational data mining (EDM) is a community interested in how to use this data to measure or optimize learning¹.
Predicting student performance can be modeled as a sequence prediction problem, or matrix completion, in the flavor of collaborative filtering models usually encountered in recommender systems. In this talk, we will show how factorization machines (FMs) can encompass several existing models in the EDM literature (notably item response theory) as special cases.
We show, using several real large-scale datasets, that FMs can estimate student knowledge accurately even when the observations are sparse, and handle side information such as knowledge components or number of attempts. Our approach allows to train models of higher dimension than existing models, and provides a testbed to try new combinations of features. The question then remains: are Deep FMs better?
¹ See our upcoming workshop Optimizing Human Learning in Montréal on June 12: https://humanlearn.io