Variance inflation is caused by a mismatch between linear projections of test and training data when projections are estimated on training sets smaller than the dimensionality of the feature space. We demonstrate that variance inflation can lead to an increased neuroimage decoding error rate for Support Vector Machines. However, good generalization may be recovered in part by a simple renormalization procedure. We show that with proper renormalization, cross-validation based parameter optimization leads to the acceptance of more non-linearity in neuroimage classifiers than would have been obtained without renormalization.
Lecture Notes in Computer Science: International Workshop, Mlini 2011, Held at Nips 2011, Sierra Nevada, Spain, December 16-17, 2011, Revised Selected and Invited Contributions, 2012, p. 256-263
Support Vector Machines; Generalizability; Variance inflation; Imbalanced data
Main Research Area:
Lecture Notes in Computer Science
International Workshop on Machine Learning and Interpretation in Neuroimaging (MLINI 2011)