A Transductive Bound for the Voted Classifier with an Application to Semi-Supervised Learning
M.-R. Amini and F. Laviolette and N. Usunier Poster ID : T24 Practical Result : A bound over the transductive risk of the bayes classifier which is tight when we have an accurate approximation of the risk of the Gibbs classifier and that the Bayes classifier makes most of its errors on low margin examples.

Link with semi-supervised learning and application :

Empirical Results :
Dataset Coil2 Digit G241c USPS PIMA WDBC d 241 241 241 241 8 30 l+u 1500 1500 1500 1500 768 569 l 10 10 10 10 10 10 SLA .302 ±.042 .201 ±.038 .314 ±.037 .342 ±.024 .379 ±.026 .168 ±.016 SLA .255±.019 .149±.012 .248±.018 .278 ±.022 .305±.021 .124±.011 TSVM .286 ±.031 .156±.014 .252±.021 .261±.019 .318 ±.018 .141 ±.016 l 100 100 100 100 50 50 SLA .148 ±.015 .091 ±.01 .201 ±.017 .114 ±.012 .284 ±.019 .112 ±.011 SLA .134±.011 .071±.005 .191±.014 .112 ±.012 .266±.018 .079±.007 TSVM .152 ±.016 .087 ±.009 .196±.022 .103±.011 .276±.021 .108 ±.01