You mention that: " In our case, when computing the impurity score with respect to a potential split, we simply average the impurity scores with respect to each output."
So what you are saying is that you do not account for the covariance of outputs directly. This is somewhat account for when averaging the impurity score for each output, correct? ps: I have only skimmed through the paper you sent and may have missed something. -- Flavio On Wed, Sep 5, 2012 at 10:43 AM, Gilles Louppe <[email protected]> wrote: > Hi Flavio, > > This is similar to [1, section 2.2.2 § "Learning"]. You can also find > a complete description in our user guide [2]. > > [1]: > http://www.montefiore.ulg.ac.be/services/stochastic/pubs/2009/DMWG09/dumont-visapp09-shortpaper.pdf > [2]: http://scikit-learn.org/dev/modules/tree.html#multi-output-problems > > I skimmed through the paper you mention. If I am correct the setting > is the same, but the impurity score is not. In our case, when > computing the impurity score with respect to a potential split, we > simply average the impurity scores with respect to each output. > > Hope this helps. > > Gilles > > On 5 September 2012 15:18, Flavio Vinicius <[email protected]> wrote: >> Hello all, >> >> I just read the release announcement, congratulations! One new caught >> my attention was: Regression Trees/Forests which support multiple >> outputs. Can someone point out any reference (papers) which this >> implementation was based on? >> >> For a while in the past I experimented with the Multivariate random >> forest which is described here: >> http://onlinelibrary.wiley.com/doi/10.1002/widm.12/abstract. >> >> Basically, the idea is that when choosing the feature to split the >> algorithm accounts for the covariance matrix of response variables*. I >> would like to know if if sklearn's implementation follow's the same >> approach. >> >> *For example, the squared Mahalanobis distance could be used for >> regression. This is the same as squared error with one output only. >> >> -- >> Flavio >> >> ------------------------------------------------------------------------------ >> Live Security Virtual Conference >> Exclusive live event will cover all the ways today's security and >> threat landscape has changed and how IT managers can respond. Discussions >> will include endpoint security, mobile security and the latest in malware >> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ >> _______________________________________________ >> Scikit-learn-general mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general > > ------------------------------------------------------------------------------ > Live Security Virtual Conference > Exclusive live event will cover all the ways today's security and > threat landscape has changed and how IT managers can respond. Discussions > will include endpoint security, mobile security and the latest in malware > threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ > _______________________________________________ > Scikit-learn-general mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general ------------------------------------------------------------------------------ Live Security Virtual Conference Exclusive live event will cover all the ways today's security and threat landscape has changed and how IT managers can respond. Discussions will include endpoint security, mobile security and the latest in malware threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
