The FAST toolkit for Unsupervised Learning of HMMs with Features
Yun Huang, José P. González-Brenes, Peter L. Brusilovsky · D-Scholarship@Pitt (University of Pittsburgh) · 2015
FAST, is an toolkit for adding features to Hidden Markov Models (HMM). It implements a recent variation of the Expectation-Maximization algorithm (Berg-Kirkpatrick et al, 2010) that allows to use logistic regression in unsupervised learning. We demonstrate FAST for predicting future student performance. Our toolkit is up to 300x faster than BNT (a Bayesian Network toolkit), and up to 25% better than conventional HMMs (with no features).