hi all, my feeling is that new SGD schemes (Averaged SGD and recent efforts in online learning) would be a nice addition. There is also an open PR on ranking with SGD using a pairwise hinge loss.
Alex On Fri, Mar 22, 2013 at 1:23 PM, Andreas Mueller <[email protected]> wrote: > Hi Anne. > Thanks for the offer. > I'm not sure we want a newtons method implementation. There is on in > liblinear. but that is one-vs-rest. > If we start reimplementing parts of liblinear, we might open pandoras box ;) > In principal I could imagine a "MultinomialLogisticRegression" estimator. > The speed should be comparable with LinearSVC, though, > which might not be that easy. > > Currently, an SGD implementation would be great. > > Cheers, > Andy > > > On 03/22/2013 01:17 PM, Anne Dwyer wrote: > > Andy, > > I wrote Python code for Newton's method logistic regression and a plot of > the hyperplane. Is this something the GSoC project would be interested in or > is it too low level? > > Anne Dwyer > > On Fri, Mar 22, 2013 at 6:58 AM, Andreas Mueller <[email protected]> > wrote: >> >> Hi Ricardo. >> I think you forgot to mention what [1] and [2] are. >> What is the difference between a relative neighborhood graph and a >> neighborhood graph? >> >> To me that sounds a bit to special purpose for the moment. >> We need Logistic Regression first (which might also be a good GSoC >> project)! >> >> Just my opinion though ;) >> >> Cheers, >> Andy >> >> >> >> On 03/22/2013 06:49 AM, Ricardo Corral C. wrote: >> >> Ok, this is a brief description of what I'm interested in. >> >> Recently, I faced a problem of evaluating the quality of a method to >> obtain features from protein structures. >> I adopted the approach given in [1] to measure separability of my >> classes independently of my capacity of make good predictions. >> This is basically a hypothesis testing of whether or not the >> distribution of classes over feature vectors is somewhat random. >> This test is made over the construction of a Relative Neighbourhood >> Graph, which is O(n^3), thus, so prohibitive for practical use. >> There is an efficient method for constructing RNG on the plane >> described in [2] O(n*log(n)), but O(n^2) for a higher d dimension (in >> fact O(n^2*f(d)) with f(d) <= (2*sqrt(d) +2)^d...). >> >> Actually, I have the test implemented, and I'm refining a speedup of >> RNG construction based on the Half-Space Proximal (HSP) graph. This is >> O(n^2log(n)), and there is no dependence of dimension other than time >> consumed in calculating distances. >> >> This is made by doing RNG test over edges in HSP (attached images for >> clarify this). >> >> Could this be of interest for sklearn users? And if so, be considered for >> GSoC? >> >> >> On Thu, Mar 21, 2013 at 12:02 PM, Andreas Mueller >> <[email protected]> wrote: >> >> On 03/21/2013 06:56 PM, Ricardo Corral C. wrote: >> >> I would like to contribute with an idea different from those listed. >> Is this the place to describe my proposal? >> >> >> I think posting it on the mailing list (at least a short description) >> would be a good start. >> Also starting to contribute ;) >> >> >> ------------------------------------------------------------------------------ >> Everyone hates slow websites. So do we. >> Make your web apps faster with AppDynamics >> Download AppDynamics Lite for free today: >> http://p.sf.net/sfu/appdyn_d2d_mar >> _______________________________________________ >> Scikit-learn-general mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general >> >> >> >> >> ------------------------------------------------------------------------------ >> Everyone hates slow websites. So do we. >> Make your web apps faster with AppDynamics >> Download AppDynamics Lite for free today: >> http://p.sf.net/sfu/appdyn_d2d_mar >> >> >> >> _______________________________________________ >> Scikit-learn-general mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general >> >> >> >> >> ------------------------------------------------------------------------------ >> Everyone hates slow websites. So do we. >> Make your web apps faster with AppDynamics >> Download AppDynamics Lite for free today: >> http://p.sf.net/sfu/appdyn_d2d_mar >> _______________________________________________ >> Scikit-learn-general mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general >> > > > > ------------------------------------------------------------------------------ > Everyone hates slow websites. So do we. > Make your web apps faster with AppDynamics > Download AppDynamics Lite for free today: > http://p.sf.net/sfu/appdyn_d2d_mar > > > > _______________________________________________ > Scikit-learn-general mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general > > > > ------------------------------------------------------------------------------ > Everyone hates slow websites. So do we. > Make your web apps faster with AppDynamics > Download AppDynamics Lite for free today: > http://p.sf.net/sfu/appdyn_d2d_mar > _______________________________________________ > Scikit-learn-general mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general > ------------------------------------------------------------------------------ Everyone hates slow websites. So do we. Make your web apps faster with AppDynamics Download AppDynamics Lite for free today: http://p.sf.net/sfu/appdyn_d2d_mar _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
