🌐 Mailman 3 May 2017 - scikit-learn - python.org

develop install with pip?
by Albert Thomas May 31, 2017

May 31, 2017

Hi all, For a develop install it is suggested in the contributing section of the website http://scikit-learn.org/stable/developers/contributing.html to do: python setup.py develop However I read on stackoverflow that the preferred way to do this is now to use pip instead of using setuptools directly: pip install -e . Thanks, Albert

2 2

KNearestNeighbour is not running in multithread
by LUONGO, ALESSANDRO May 30, 2017

May 30, 2017

Hi everyone! I'm successfully using scikit-learn on a 384 core machine. I'm playing with two deployment: The first is a anaconda installation of python, which use MKL as backend of numpy, with python 3.6 The second is a "native" installation of scikit-learn and numpy, and thus the backend is based on openblas and python 3.4.5 Both implementations works, and I can see a high number of threads wigh high CPU load. (for instance when I'm doing PCA) The problem that I don't know how to debug, is that with kNearestNeighbour is using only one core. This puzzle me, since I can see that since version 0.17, the PR with the parallel KNN has been accepted into the main branch. https://github.com/scikit-learn/scikit-learn/pull/4009 , Sklearn should have merged this changes 1 year ago, and my version of sklearn is: > print('The scikit-learn version is {}.'.format(sklearn.__version__)) > The scikit-learn version is 0.18.1. Do you have any hints on how to use parallel KNN? I'm classifying a high dimensional dataset of MNIST (image digits). So I'm doing PCA to get vector of dimension 35-50, and then I'm doing a nonlinear expansion, so I'm getting vector of dimension 600-100. That's why I need parallelism so badly. clf = KNeighborsClassifier(algorithm='ball_tree') clf = clf.fit(train, train_labels) Thanks for all your amazing work. Ale

3 2

Failing check_estimator on py-earth
by Jason Rudy May 20, 2017

May 20, 2017

I'm pushing to get py-earth ready for a release, but I'm having an issue with the check_estimator function on 32 bit windows machines. Here is a link to the failing build on appveyor: https://ci.appveyor.com/project/jcrudy/py-earth/build/job/21r6838yh1bgwxw4 It appears that array conversion is producing some small differences that make check_estimators_data_not_an_array fail. I'll probably have to set up a 32 bit environment with a debugger and drill down to find the bug, but I'm wondering if anybody here has tips or experience that might help me guess the problem without doing that. I am pretty ignorant about numpy type standards and conversions, so even something that seems obvious to you might help me. Best, Jason

3 4

Re: [scikit-learn] scikit-learn Digest, Vol 14, Issue 6
by Mamun Rashid May 19, 2017

May 19, 2017

Hi J.B and the list. Please accept my apology for a much delayed response. Was ill for last few days and did not access my email. Thanks for your detailed response. > What does "performing better" mean in this case? How are you defining performance? A particular metric such as MCC, PPV, or NPV? I was looking at precision recall. I have a huge class imbalance [positive class is much smaller than negative class]. So, I am testing performance of various classifiers with an increasing negative set size ( every time I am randomly selecting a larger negative set ). It seems SVC shows better performance in Precision recall space ( SVC precision recall curve is above RFC curve ). Because of the two following issues : 1. I have a major class imbalance 2. Some of my positive observations are sometimes tightly packed within negative observation clusters [ Observations from 2 dimensional PCA and tSNE plot ]. My aim is to obtain a very clean set of positive predictions as a trade-off I am happy to sacrifice some of the positive observations > Also, how is the cross-validation being done - is the data shuffled before creating train/test groups are created? Is the exact same split of training and test data per fold used for both > SVC and RF? I am currently testing it. Thanks for the suggestion. > Normalizing your continuous values seems quite fine, but consider these > aspects: > --Does it make sense in the domain of your problem to Z-normalize the integral (integer-valued) descriptors/features? > For the integral values, would subtracting about the median value make more sense? This is similar to the previous consideration. Yes. Z-score normalisation does not make much sense. Thanks for pointing it out. Currently testing it. > --What happens to SVC if you don't normalise? SVC performs quite badly. > --What happens to RF if you do normalise? This is interesting. My understating was that decision tree based algorithms does not require normalised data. I took your suggestion and tested an RFC with and without normalised data. Their result [Confusion matrix at 0.5 operating point] seems to be identical. It felt odd to me. I have only tested on a small data set. Currently running it on different data sets to see if this is persistent. Would you have expected this ? > If you can provide a good amount of concrete data to present along with your "problem", this community is excellent at providing intelligent, helpful responses. Thanks a lot for your suggestion. I will try to create some example data sets and results from the current analysis and post it as soon as possible. Thanks in advance for your help. Regards, Mamun > Today's Topics: > > 1. SVC data normalisation (Mamun Rashid) > > Message: 2 > Date: Mon, 8 May 2017 21:48:28 +0900 > From: "Brown J.B." <jbbrown(a)kuhp.kyoto-u.ac.jp> > > Dear Mamun, > > *A.* 80% features are binary [ 0 or 1 ] >> *B.* 10% are integer values representing counts / occurrences. >> *C.* 10% are continuous values between different ranges. >> >> My prior understanding was that decision tree based algorithms work better >> on mixed data types. In this particular case I am noticing >> SVC is performing much better than Random forest. >> > > What does "performing better" mean in this case? > How are you defining performance? > A particular metric such as MCC, PPV, or NPV? > > Also, how is the cross-validation being done - is the data shuffled before > creating train/test groups are created? > Is the exact same split of training and test data per fold used for both > SVC and RF? > > >> I Z-score normalise the data before I sent it to support vector >> classifier. >> - Binary features ( type *A) *are left as it it. >> - Integer and Continuous features are Z-score normalised [ ( feat - >> mean(feat) ) / sd(feat) ) . >> > > Normalizing your continuous values seems quite fine, but consider these > aspects: > --Does it make sense in the domain of your problem to Z-normalize the > integral (integer-valued) descriptors/features? > --For the integral values, would subtracting about the median value make > more sense? This is similar to the previous consideration. > --What happens to SVC if you don't normalize? > --What happens to RF if you do normalize? > > While my various comments above are all geared toward empirical aspects and > not toward theoretical aspects, picking some of them to explore is likely > to help you gain practical insight on your situation/inquiry. > I'm sure you already know this, but while machine learning may have some > "practical guidelines for best practices", they are guidelines and not hard > rules. > So, again, I would recommend doing some more empirical tests and > re-evaluating your situation once you have new data in hand. > > If you can provide a good amount of concrete data to present along with > your "problem", this community is excellent at providing intelligent, > helpful responses. > > Hope this helps. > > J.B. >

1 0

impurity criterion in gradient boosted regression trees
by Olga Lyashevska May 11, 2017

May 11, 2017

Hi all, I am trying to understand differences in feature importance plots obtained with R package gbm and sklearn. Having compared both implementation side by side it seems that the models are fairly similar, however feature importance plots are rather distinct. R uses empirical improvement in squared error as it is described in Friedmans's "Greedy Function Approximation" paper (eq. 44, 45). sklearn (as far as I could see it in the code) uses the weighted reduction in node purity. How exactly is this calculated? Is it a gini index? Is there a reference? I found this, but I find this hard to follow: https://github.com/scikit-learn/scikit-learn/blob/fc2f24927fc37d7e42917369f… I have also seen a post by Matthew Drury on stack exchange: https://stats.stackexchange.com/questions/162162/relative-variable-importan… Many thanks, Olga

2 1

Broken c dependencies
by Georg Heiler May 10, 2017

May 10, 2017

Hi, unfortunately, the c dependencies of my scikit-learn installation broke and I get the following error on osx: dlopen(/usr/local/lib/python3.6/site-packages/sklearn/svm/libsvm.cpython-36m-darwin.so, 2): Symbol not found: __ZdlPvm Referenced from: /usr/local/lib/python3.6/site-packages/sklearn/svm/libsvm.cpython-36m-darwin.so (which was built for Mac OS X 10.12) Expected in: /usr/lib/libstdc++.6.dylib in /usr/local/lib/python3.6/site-packages/sklearn/svm/libsvm.cpython-36m-darwin.so Even removing my python installation and re-installing does not seem to get this library back. Regards, Georg

5 11

SVC data normalisation
by Mamun Rashid May 8, 2017

May 8, 2017

Hi All, I am testing two classifiers [ 1. Random forest 2. SVC with radial basis kernel ] on a data set via 5 fold cross validation. The feature matrix contains : A. 80% features are binary [ 0 or 1 ] B. 10% are integer values representing counts / occurrences. C. 10% are continuous values between different ranges. My prior understanding was that decision tree based algorithms work better on mixed data types. In this particular case I am noticing SVC is performing much better than Random forest. I Z-score normalise the data before I sent it to support vector classifier. - Binary features ( type A) are left as it it. - Integer and Continuous features are Z-score normalised [ ( feat - mean(feat) ) / sd(feat) ) . I was wondering if anyone can tell me if this normalisation approach it correct for SVC run. Thanks in advance for your help. Regards, Mamun

2 1

using a mask for brain images
by Alle Meije Wink May 5, 2017

May 5, 2017

I have a script to classify MRI perfusion maps from healthy subjects and patients. For the file IO and the classifier I have started with the example code in Abraham et al 2014 [https://arxiv.org/pdf/1412.3919.pdf]. I use the same classifier as in the paper to produce a back-projected map of classification weights, which I then want to 'unmask' like in the paper: coef=clf.coef_ coef=featureselection.inverse_transform(coef) and map_name='weights_check.nii.gz' wmap=np.zeros(mask.shape, dtype=X.dtype) wmap[mask]=coef img=nb.Nifti1Image(wmap,np.eye(4)) img.to_filename(map_name) But the line "wmap[mask]=coef" throws an error "ValueError: boolean index array should have 1 dimension". I tried the example code from the paper and that works. Is the 'coef' array of back-projected SVM weights in some way different than the masked input image? Or am I doing something else wrong? The error suggests that the mask array is the problem. The complete script is attached. Many thanks for your help! Alle Meije

2 3

Contribution to sklearn: Cross validation of time series
by andres lago May 5, 2017

May 5, 2017

Hello, I'd like to contribute with a new functionality in sklearn. It's the cross validation of time series. It's an evolution of the current functionality, implemented by TimeSeriesSplit. TimeSeriesSplit only allows the user to set the number of folds. In real life, when performing the cross validation of time series, other parameters are required, for instance: -minimum size of CV-training set -size of CV-test set -fixed or variable length of CV-training set. The functionality is inspired by the R library 'caret'. If you agree, I can share my code. I developed it for a project with the french rail company SNCF. It's in production now. Regards, Andres

3 5

question about scikit-learn
by 熊瑶 May 5, 2017

May 5, 2017

Dear professor, scikit-learn is really good for me to do some work using machine learning method. Here, I have two questions: 1）To do 5 fold cross-validation, when I use StratifiedKFold，I could get stratified folds that each fold contains approximately the same percentage of samples of each target class as the complete set. And, when I use GroupKFold, it ensures that the same group is not represented in both testing and training sets. I want to know whether there is a method to combine these two methods together? 2) When I use GridSearchCV to do parameter search, I use scoring="accuracy" as scoring function to choose the best parameters. And I find that I can only get the accuracy score from the 5 fold cross-validation. What can I do if I want to get other scores such as sensitivity, specificity, MCC at the same time? It means that I want to use accuracy to choose the best parameters and I want to get the scores of many scoring parameters at the same time when I do 5 fold cross-validation. Thank you. 熊瑶北京大学深圳研究生院化学生物学与生物技术学院 XIONG Yao G301, School of Chemical Biology & Biotechnology Peking University Shenzhen Graduate School Shenzhen 518055, Guangdong, P.R. China E-mail: xiongyao(a)pku.edu.cn or xiongy20121126(a)foxmail.com

2 1