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

Accessing Clustering Feature Tree in Birch
by Sema Atasever Sept. 13, 2017

Sept. 13, 2017

Dear scikit-learn members, Considering the "CF-tree" data structure : - How can i *access Clustering Feature Tree* in Birch? - For example, how many clusters are there in the hierarchy under the *root node* and what are the data samples in this cluster? - Can I get them separately for 3 trees? Best.

3 3

FW: Random Forest Regressor criterion
by Evans J.R.A. Aug. 31, 2017

Aug. 31, 2017

Hi there, I would like to fully understand how the Random Forest Regressor chooses how to split the data at each node. I understand that each tree considers a boostrap sample of the training data, and on each split a random subset of features (using max_features) are considered. But among these features, how does the algorithm work out which is the best split to make? I am using the default criterion 'mse', but don't understand the given explanation "equal to variance reduction as feature selection criterion". Does this mean that for each possible split that could be made, the sum of variances of data in the child nodes is calculated, then the algorithm would use the split with the least sum of variances? Kind regards, Jonny Evans Doctoral Researcher Transportation Research Group Faculty of Engineering and the Environment University of Southampton Email: Jonny.Evans(a)soton.ac.uk<mailto:Jonny.Evans@soton.ac.uk>

2 1

Getting weight coefficient of logistic regression from a pipeline
by Raga Markely Aug. 29, 2017

Aug. 29, 2017

Hello, I am wondering if it's possible to get the weight coefficients of logistic regression from a pipeline? For instance, I have the followings: > clf_lr = LogisticRegression(penalty='l1', C=0.1) > pipe_lr = Pipeline([['sc', StandardScaler()], ['clf', clf_lr]]) > pipe_lr.fit(X, y) Does pipe_lr have an attribute that I can call to get the weight coefficient? Or do I have to get it from the classifier as follows? > X_std = StandardScaler().fit_transform(X) > clf_lr = LogisticRegression(penalty='l1', C=0.1) > clf_lr.fit(X_std, y) > clf_lr.coef_ Thank you, Raga

3 10

scikit-learn-commits mailing list defunct?
by Andreas Mueller Aug. 28, 2017

Aug. 28, 2017

Hey all. Is it just me or is the scikit-learn-commits mailing list no longer working? Given that it's still on sourceforge, that seems somewhat likely. I find the mailing list helpful in case I can't keep track of the issue tracker (i.e. for the last 3 years?). It looks like it was set up here: https://github.com/scikit-learn/scikit-learn/settings/hooks/28838 I propose we transition this to either a google group or another python.org mailing list (they already gave us 2 ;). Cheers, Andy

2 2

remoe from list
by Liu, Susan Aug. 28, 2017

Aug. 28, 2017

hi there, just wanted to ask if i could be removed from list? Thanks, Susan

2 1

imbalanced-learn 0.3.0 is chasing scikit-learn 0.19.0
by Guillaume Lemaître Aug. 28, 2017

Aug. 28, 2017

We are excited to announce the new release of the scikit-learn-contrib imbalanced-learn, already available through conda and pip (cf. the installation page https://tinyurl.com/y92flbab for more info) Notable add-ons are: * Support of sparse matrices * Support of multi-class resampling for all methods * A new BalancedBaggingClassifier using random under-sampling chained with the scikit-learn BaggingClassifier * Creation of a didactic user guide * New API of the ratio parameter to fit the needs of multi-class resampling * Migration from nosetests to pytest You can check the full changelog at: http://contrib.scikit-learn.org/imbalanced-learn/stable/whats_new.html#vers… A big thank you to contributors to use, raise issues, and submit PRs to imblearn. -- Guillaume Lemaitre

10 9

Clarification regarding SGDClassifier
by Shabie Iqbal Aug. 27, 2017

Aug. 27, 2017

Dear all, could anyone of you kindly answer this question I posted on Stackoverflow: https://stackoverflow.com/questions/45900330/sklearn-sgdclassifiers-decisio… Best, Shabie

2 2

Decision stubs?
by Stuart Reynolds Aug. 27, 2017

Aug. 27, 2017

Is it possible to efficiently get at the branch statistics that decision tree algorithms iterate over in scikit? For example if the root population has the class counts in the output vector: c0: 5000 c1: 500 Then I'd like to iterate over: # For a boolean (2 valued category) f1=True: c0=3000, c1=450 f1=False: c0=300, c1=30 f1=Null: c0=1700, c1=20 # ? Is considered? # For a continuous value f2<10: c0= ... c1= ... f2>=10: c0= ... c1= ... f2<22: c0= ... c1= ... f2>=22: c0= ... c1= ... I'd like to experiment with building models on-demand for each input row in a predict. To work efficiently, I'd like to reduce the training set to the 'most significant' sub-space(s) using the population statistics. I can do it in pandas, although its fairly inefficient to iterate over each feature column many times. Thanks, - Stu

1 0

L-BFGS in MLPClassifier
by Dr. Mario Michael Krell Aug. 26, 2017

Aug. 26, 2017

To whoever programmed the MLPClassifier (with the L-BFGS solver), I just wanted to personally thank you and if I get your name(s), I would mention it/them in my paper additionally to the mandatory sklearn citation. I hope that sklearn will be keeping this algorithm forever in their library despite the increasing amount of established deep learning libraries that seem to make this code obsolete. For my small scale, more theoretic analysis, it worked much better than any other algorithm and I would not have gotten such surprising results. Due to the high quality implementation, the integration of a much better solver than SGD, and the respective good documentation, I could show empirically how the VC dimension and another property of MLPs (MacKay dimension) actually scale linear with the number of edges in the respective graph which helped us to provide a new much more strict upper bound (https://arxiv.org/abs/1708.06019 <https://arxiv.org/abs/1708.06019>). This would have not been possible with other implementations. If there is an interest by the developers, I could try to contribute a tutorial documentation for sklearn. Just let me know. Thank you a lot!!! Best, Mario

2 1

different sized inputs in call to custom metric in KNN
by Stephen O'Neill Aug. 25, 2017

Aug. 25, 2017

Hey Gang, I was wondering if anyone might be able to answer a question about the sklearn,neighbors.NearestNeighbors class. For reference, I'm on: anaconda distribution python 2.7.11 sklearn version 0.17.1 I'm subclassing the NearestNeighbors class and using a custom distance metric, something like the following: class MyModel(NearestNeighbors): def __init__(self, some_info): def custom_dist(x, y, info=some_info): return numpy.sum(numpy.abs(x - y)/some_info) return scalar_value NearestNeighbors.__init__(self, metric=custom_dist) So I build a dummy dataset based on some gaussians of shape (5000,3), then later when I call MyModel.fit() I get the error: "ValueError: operands could not be broadcast together with shapes (10,) (3,)" inside of my custom_dist function. Naturally I checked with some simple print x, print y statements inside of custom_dist, and sure enough the shapes of x and y are both (10,), whereas I am expecting them to be of shape (3,) since my dummy data has 3 columns. (Note the actual custom_dist function written above is not what I'm truly using but it does reproduce the same ValueError). When I change my NearestNeighbors.__init__ call to (self, algorithm='brute') instead of the default algorithm='auto', the x and y values that get passed to my custom_dist are shape (3,) like I would expect. What is different in the distance metric between the algorithm='auto' and algorithm='brute' that would transform a 3-dimensional sample to a 10-dimensional sample? Do the KDTree and/or BallTree classes use the distance metric on tree nodes or something too? I wasn't able to figure out where the shape (10,) x and y samples could be coming from. Thanks in advance! Best, Steve O'Neill

1 0