2013/4/26 Eustache DIEMERT <eusta...@diemert.fr>: > I tried a very naive bootstrapping approach (learn RF on supervised data, > predict on unsupervised, then learn RF on all) but with no luck either.
Sounds like you've now got a self-training algorithm with only one iteration. You may have more luck with a proper self-training/EM algorithm, e.g. DCEM [1] (works sometimes, esp. with the error modelling) or CVEM [2] (not tried that one yet). Such algorithms take the classifier's confidence into account when doing the next fit, by doing something like class_weight=predict_proba. You can also try learning extra features with k-means [3], or just try the LabelPropagation algorithm in scikit-learn. [1] http://ama.liglab.fr/~amini/Publis/SemiSupImpSpr.pdf [2] http://research.microsoft.com/en-us/um/people/xiaohe/nips08/paperaccepted/nips2008wsl1_02.pdf [3] http://fastml.com/the-secret-of-the-big-guys/ -- Lars Buitinck Scientific programmer, ILPS University of Amsterdam ------------------------------------------------------------------------------ Try New Relic Now & We'll Send You this Cool Shirt New Relic is the only SaaS-based application performance monitoring service that delivers powerful full stack analytics. Optimize and monitor your browser, app, & servers with just a few lines of code. Try New Relic and get this awesome Nerd Life shirt! http://p.sf.net/sfu/newrelic_d2d_apr _______________________________________________ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
