Mailman 3 April 2017 - scikit-learn

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

gridsearchCV able to handle list of input?
by Carlton Banks May 1, 2017

May 1, 2017

I am currently trying to run some gridsearchCV on a keras model which has multiple inputs. The inputs is stored in a list in which each entry in the list is a input for a specific channel. Here is my model and how i use the gridsearch. https://pastebin.com/GMKH1L80 <https://pastebin.com/GMKH1L80> The error i am getting is: https://pastebin.com/A3cB0rMv <https://pastebin.com/A3cB0rMv> Any idea how i can resolve this?

3 13

RFE/RFECV parameter suggestion
by George Fisher April 30, 2017

April 30, 2017

I found that xgboost generates an exception under RFECV when the number of features remaining falls below 3. I fixed this for myself by adding a 'stop_at' parameter (default=1) that stops the process in RFE when the remaining features falls below this number. I think it might be a useful feature more broadly than simply as a hacked work-around so I offer it as a pull request. George Fisher george(a)georgefisher.com +1 917-514-8204 https://github.com/grfiv Ubuntu 17.04 Desktop Python 3.5.3 IPython 6.0.0 sklearn 0.18.1 (xgboost 0.6)

3 2

Plotting Code in Examples Pull Request
by Isaac Laughlin April 26, 2017

April 26, 2017

I'm hoping to get some more comments on this pull request: https://github.com/scikit-learn/scikit-learn/pull/8490 I've got some availability and would love to get this resolved and start working on some others if possible. Thanks in advance!

1 0

Ordinary Least Square Regression Under-determined system.
by Joshua Mannheimer April 25, 2017

April 25, 2017

Hi all, So I am trying to write a Principle Components Regression implementation in Python to match the PLS package in R. I am getting better results in R so I am trying to figure out where the discrepancy was. The data I am using is way undetermined where n_features ~ 50,000 and n_samples ~ 500 thus why PCR is necessary. Just to see what would happen I used sklearn.LinearRegression on the original 500 X 50000 dataset. I expected I would get an error message stating the the system was not solvable but it worked and I got an answer that was at least on par with the PCR solution. So I am wondering how it is possibly solving this system if anybody knows. Thanks -- Joshua D. Mannheimer M.E. Biomedical Engineering Ph.d Student Flint Animal Cancer Research Center Office: A 259 CSU Veterinary Campus Colorado State University (970)-389-3951 jmannhei(a)rams.colostate.edu

3 2

Scikit-learn sprint @Paris 6-10 June
by Guillaume Lemaitre April 24, 2017

April 24, 2017

Dear all, As previously mentioned, the scikit-learn community will gather for a coding sprint in Paris from the 6th to the 10th of June, ahead of the PyParis conference. All information regarding this event is available on the following wiki page: https://github.com/scikit-learn/scikit-learn/wiki/Upcoming-events Do no hesitate to consult this page regularly; any new information will be published there. This event is also opened to new contributors which can refer to the above link. Any participant will have to add his presence by editing the wiki page. Hope to see you soon in Paris. -- Guillaume Lemaitre INRIA Saclay - Parietal team Center for Data Science Paris-Saclay https://glemaitre.github.io/

1 0

What if I don't want performance measures per each outcome class?
by Suranga Kasthurirathne April 24, 2017

April 24, 2017

Hello all, I'm looking at the confidence matrix and performance measures (precision, recall, f-measure etc.) produced by scikit. It seems that scikit calculates these measures per each outcome class, and then combines them into some sort of average. I would really like to see these measures presented in the traditional(?) context, where sensitivity is TP / TP + FN. (and is combined, and NOT per class!) If I were to take scikit predictions, and calculate sensitivity using the above, then my results wont match up to what scikit says :( How can I switch to seeing overall performance measures, and not per class? and also, how may I obtain 95% confidence intervals foreach of these measures? -- Best Regards, Suranga

2 1

sklearn - knn sklearn.neighbors kneighbors function producing unexpected result for text analysis?
by Evaristo Caraballo April 21, 2017

April 21, 2017

I have been asked to implement a simple knn for text similarity analysis. I tried by using sklearn.neighbors module.The file to be analysed consisted on 2 relevant columns: "text" and "name".The knn model should be fitted with bag-of-words of a corpus of around 60,000 pre-treated text fragments of about 200 words each. I used CounterVectorizer.As test I was asked to use the model to get the names in the "name" column related to 10 top text strings that are the closest to a pre-selected one that also exists in the corpus used to initialise the knn model. Similarity distance should be measured using an euclidean metric.I used the kneighbors function to obtain the closest neighbors.Below you can find the code I was trying to implement using kneighbors:import os, sys import sklearn import sklearn.neighbors as sk_neighbors from sklearn.feature_extraction.text import CountVectorizer import pandas import scipy import matplotlib.pyplot as plt import numpy as np %matplotlib inline wiki = pandas.read_csv('wiki_filefragment.csv') mod_count_vect = CountVectorizer() count_vect = mod_count_vect.fit_transform(wiki['text']) print(count_vect.shape) mod_count_vect.get_feature_names() mod_enc = sklearn.preprocessing.LabelEncoder().fit(wiki['name']) enc = mod_enc.transform(wiki['name']) enc model = sk_neighbors.NearestNeighbors( n_neighbors=10, algorithm='brute', p = 2 ) #no matter what I use, it is always the same modelfit = model.fit(count_vect, enc) #also likely the kneighbors is not working? print( mod_enc.inverse_transform( modelfit.kneighbors( count_vect[mod_enc.transform( ['Franz Rottensteiner'] )], n_neighbors=11, return_distance=False ) ) )This implementation gave me the following results for the first 10 nearest neighbors to 'Franz Rottensteiner': Franz Rottensteiner, Ren%C3%A9 Froger, Ichikawa Ennosuke III, Tofusquirrel , M. G. Sheftall, Peter Maurer, Allan Weisbecker, Ferdinand Knobloch, Andrea Foulkes, Alan W. Meerow, John Warner (writer) The results continued to be far from being close to the test solution (which use Graphlab Create and SFrame), which are: Franz Rottensteiner, Ian Mitchell (author), Rajiva Wijesinha, Andr%C3%A9 Hurst, Leslie R. Landrum, Andrew Pinsent, Alan W. Meerow, John Angus Campbell, Antonello Bonci, Henkjan Honing, Joseph Born Kadane In fact, I tried a simple brute force implementation by iterating over the list of texts calculating distances with scipy and that gave me the expected results. The result was the same after also using Python 2.7.A link to the implementations (the one that doesn't work and the one that does) together a pick the file used for this test can be found on this Gist.Does anyone can suggest what it is wrong with my sklearn implementation?Relevant resources are: - Anaconda Python3.5 (with a virtenv using 2.7) - Jupyter - sklearn 0.18 - pandas

3 4

feature importance calculation in gradient boosting
by Olga Lyashevska April 20, 2017

April 20, 2017

Hi, I would like to understand how feature importances are calculated in gradient boosting regression. I know that these are the relevant functions: https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/ensemble/g… https://github.com/scikit-learn/scikit-learn/blob/fc2f24927fc37d7e42917369f… From the literature and elsewhere I understand that Gini impurity is calculated. What is this exactly and how does it relate to 'gain' vs 'frequency' implemented in XGBoost? http://xgboost.readthedocs.io/en/latest/R-package/discoverYourData.html My problem is that when I fit exactly same model in sklearn and gbm (R package) I get different variable importance plots. One of the variables which was generated randomly (keeping all other variables real) appears to be very important in sklearn and very unimportant in gbm. How is this possible that completely random variable gets the highest importance? Many thanks, Olga

2 2

XGboost Classifier error
by Startup Hire April 20, 2017

April 20, 2017

Hi!, I am trying to use XGBoost Classifer in RandomizedSearchCV as follows: clf = xgb.XGBClassifier() random_search_sg = RandomizedSearchCV(clf, param_distributions=params_dist, n_iter=n_iter_search, scoring=kappa_scorer, verbose=3, error_score=-1, fit_params=fit_params, n_jobs=-1) start = time() random_search_sg.fit(scaled_data, a_l) scaled_data = (0, 0) 4.53937223364 (0, 1) 4.08089927979 (0, 2) 5.08534158523 (0, 3) 0.900022077306 (0, 4) 0.582895703409 (0, 5) 3.52674131829 (0, 6) 2.00912587286 (0, 8) 1.06039501135 (0, 9) 4.8956331357 (0, 11) 1.51595206264 (0, 13) 3.00108387862 (0, 14) 0.0 (1, 0) 1.51312407788 (1, 1) 1.36029975993 (1, 2) 2.54267079261 (1, 3) 1.36638272336 (1, 4) 0.0225891281189 (1, 5) 3.52674131829 a_l = [1 0 0 ..., 0 0 0] (after using ravel) I am getting the error Python int too large to convert to C long <http://stackoverflow.com/questions/22114088/overflowerror-python-int-too-la…> while fitting the data using random_search_sg How to resolve this? Is this related to the formats of scaled data and a_l ? Regards, Sanant

2 2