2013/11/29 abdalrahman eweiwi <[email protected]>: > Well, Kernel partial least squares can be used for regression as well as for > classification. It has been shown to work as good as Gaussian processes for > regression on head pose estimation applications. See for example this CVPR > paper > > http://iselab.cvc.uab.es/files/Publications/2012/PDF/AGD2012.pdf > > It can be used as a classification machine which I frequently use. See this > paper: > > http://enpub.fulton.asu.edu/cseml/06summer/pls_dis.pdf
Interesting, thanks. > if you would like a brief report on its performance, I can run experiments > on the datasets you like and provide the results on its performance as a > classification machine, for both PLS and KPLS. It would be great to bench it quickly on the digits dataset from sklearn. It's a toy dataset that is not completely linearly separable. Please compare training speed and test accuracy using a 5 fold cross validation vs LinearSVC and SVC with a gaussian kernel with tuned values for C and gamma (assuming a rbf kernel for the SVC and kernel PLS model). You can also try to adapt the covertype benchmark to run your implementation of kernel PLS + linear svc on it: https://github.com/scikit-learn/scikit-learn/blob/master/benchmarks/bench_covertype.py But I am not sure that a kernel method can work on a dataset with that many samples. > For regression, I cant think right now of a benchmark for that, however, I > would be happy if you suggest any. That would be great indeed. In scikit-learn there is the boston house price regression dataset. Linear models do not perform as good as non-linear tree based models such as GBRT on this data. Before starting to write code for making kernel PLS part of the sklearn code base, please review the existing source code from the PLS estimator, the kernel PCA estimator and this PR to add kernel ridge: https://github.com/scikit-learn/scikit-learn/pull/2165 There might be common code to reuse and / or refactor. Also make sure you read: http://scikit-learn.org/stable/developers/ -- Olivier ------------------------------------------------------------------------------ Rapidly troubleshoot problems before they affect your business. Most IT organizations don't have a clear picture of how application performance affects their revenue. With AppDynamics, you get 100% visibility into your Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro! http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
