Liuliu writeup of Cohen and Carvalho

This is a review of Cohen_2005_stacked_sequential_learning by user:Liuliu.

This paper is about using stacked sequential learning method to solve sequential partitioning tasks, which is a sequential classification task that characterized by long runs of identical labels. It is a meta learning method that works on both non-sequential based learners and sequential based learners. When it works on MEMM, it helps solve the mismatch problem of MEMM, which is that the MEMM model trained on training data doesn’t match the test data. Because the MEMM trained model gives higher weights to history features (labels of near nodes). However, these history features in the training data are true labels while in the testing data they are predicted ones. To make the training and testing consistent, stacking method also use the predicted labels as history features instead of true labels by cross-validation.

Regarding the question about naive Bayes, it should be 12(10+2) times slower.