MEG Decoding with Hierarchical Combination of Logistic Regression and Random Forests

This document describes the solution of the second place team in the DecMeg2014 brain decoding competition hosted at Kaggle.com. The model is a hierarchical combination of logistic regression and random forest. The first layer consists of a collection of 337 logistic regression classifiers, each using data either from a single sensor (31 features) or data from a single time point (306 features). The resulting probability estimates are fed to a 1000-tree random forest, which makes the final decision. In order to adjust the model to an unlabeled subject, the classifier is trained iteratively: After initial training, the model is retrained with unlabeled samples in the test set using their predicted labels from first iteration.