Higher criticism thresholding: Optimal feature selection when useful features are rare and weak Export

@article{citeulike:3361217, abstract = {10.1073/pnas.0807471105 In important application fields today\^{a}€”genomics and proteomics are examples\^{a}€”selecting a small subset of useful features is crucial for success of Linear Classification Analysis. We study feature selection by thresholding of feature -scores and introduce a principle of threshold selection, based on the notion of (HC). For = 1, 2, \^{a}€¦, , let \"{I}€ denote the two-sided -value associated with the th feature -score and \"{I}€ denote the th order statistic of the collection of -values. The HC threshold is the absolute -score corresponding to the -value maximizing the HC objective (/ \^{a}ˆ’ \"{I}€)/ . We consider a rare/weak (RW) feature model, where the fraction of useful features is small and the useful features are each too weak to be of much use on their own. HC thresholding (HCT) has interesting behavior in this setting, with an intimate link between maximizing the HC objective and minimizing the error rate of the designed classifier, and very different behavior from popular threshold selection procedures such as false discovery rate thresholding (FDRT). In the most challenging RW settings, HCT uses an unconventionally low threshold; this keeps the missed-feature detection rate under better control than FDRT and yields a classifier with improved misclassification performance. Replacing cross-validated threshold selection in the popular Shrunken Centroid classifier with the computationally less expensive and simpler HCT reduces the variance of the selected threshold and the error rate of the constructed classifier. Results on standard real datasets and in asymptotic theory confirm the advantages of HCT.}, author = {Donoho, David and Jin, Jiashun}, citeulike-article-id = {3361217}, citeulike-linkout-0 = {http://dx.doi.org/10.1073/pnas.0807471105}, citeulike-linkout-1 = {http://www.pnas.org/content/105/39/14790.abstract}, citeulike-linkout-2 = {http://www.pnas.org/content/105/39/14790.full.pdf}, citeulike-linkout-3 = {http://view.ncbi.nlm.nih.gov/pubmed/18815365}, citeulike-linkout-4 = {http://www.hubmed.org/display.cgi?uids=18815365}, day = {30}, doi = {10.1073/pnas.0807471105}, journal = {Proceedings of the National Academy of Sciences}, keywords = {classification, feature-selection}, month = {September}, number = {39}, pages = {14790--14795}, posted-at = {2009-09-25 11:14:18}, priority = {2}, title = {Higher criticism thresholding: Optimal feature selection when useful features are rare and weak}, url = {http://dx.doi.org/10.1073/pnas.0807471105}, volume = {105}, year = {2008} }

TY - JOUR ID - citeulike:3361217 L3 - citeulike-article-id:3361217 N2 - 10.1073/pnas.0807471105 In important application fields today—genomics and proteomics are examples—selecting a small subset of useful features is crucial for success of Linear Classification Analysis. We study feature selection by thresholding of feature -scores and introduce a principle of threshold selection, based on the notion of (HC). For = 1, 2, …, , let π denote the two-sided -value associated with the th feature -score and π denote the th order statistic of the collection of -values. The HC threshold is the absolute -score corresponding to the -value maximizing the HC objective (/ − π)/ . We consider a rare/weak (RW) feature model, where the fraction of useful features is small and the useful features are each too weak to be of much use on their own. HC thresholding (HCT) has interesting behavior in this setting, with an intimate link between maximizing the HC objective and minimizing the error rate of the designed classifier, and very different behavior from popular threshold selection procedures such as false discovery rate thresholding (FDRT). In the most challenging RW settings, HCT uses an unconventionally low threshold; this keeps the missed-feature detection rate under better control than FDRT and yields a classifier with improved misclassification performance. Replacing cross-validated threshold selection in the popular Shrunken Centroid classifier with the computationally less expensive and simpler HCT reduces the variance of the selected threshold and the error rate of the constructed classifier. Results on standard real datasets and in asymptotic theory confirm the advantages of HCT. IS - 39 JF - Proceedings of the National Academy of Sciences EP - 14795 TI - Higher criticism thresholding: Optimal feature selection when useful features are rare and weak VL - 105 SP - 14790 KW - classification KW - feature-selection AU - Donoho, David AU - Jin, Jiashun PY - 2008/09/30/ UR - http://dx.doi.org/10.1073/pnas.0807471105 ER -