Mailman 3 March 2020 - scikit-learn

Understanding max_features parameter in RandomForestClassifier
by aditya aggarwal March 11, 2020

March 11, 2020

For RandomForestClassifier in sklearn max_features parameter gives the max no of features for split in random forest which is sqrt(n_features) as default. If m is sqrt of n, then no of combinations for DT formation is nCm. What if nCm is less than n_estimators (no of decision trees in random forest)? *example:* For n = 7, max_features is 3, so nCm is 35, meaning 35 unique combinations of features for decision trees. Now for n_estimators = 100, will the remaining 65 trees have repeated combination of features? If so, won't trees be correlated introducing bias in the answer? Thanks Aditya Aggarwal

3 3

Threshold for roc_curve in binary classification
by aditya aggarwal March 11, 2020

March 11, 2020

Hello I was going through the logic to calculate threshold to plot roc_curve. As far as I could understand, fps, tps and threshold is calculated in slklearn.metrics._binary_clf_curve . How are multiple values of threshold calculated for binary classification? Also what is happening in the following lines? distinct_value_indices = np.where(np.diff(y_score))[0] threshold_idxs = np.r_[distinct_value_indices, y_true.size - 1] Thanks

1 0

distances
by Joel Nothman March 6, 2020

March 6, 2020

I noticed a comment by @amueller on Gitter re considering a project on our distances implementations. I think there's a lot of work that can be done in unifying distances implementations... (though I'm not always sure the benefit.) I thought I would summarise some of the issues below, as I was unsure what Andy intended. As @jeremiedbb said, making n_jobs more effective would be beneficial. Reducing duplication between metrics.pairwise and neighbors._dist_metrics and kmeans would be noble (especially with regard to parameters, where scicpy.spatial's mahalanobis available through sklearn.metrics does not accept V but sklearn.neighbors does). and perhaps offer higher consistency of results and efficiencies. We also have idioms the code like "if the metric is euclidean, use squared=True where we only need a ranking, then take the squareroot" while neighbors metrics abstract this with an API by providing rdsit and rdist_to_dist. There are issues about making sure that pairwise_distances(metric='minkowski', p=2) is using the same implementation as pairwise_distances(metric='euclidean'), etc. We have issues with chunking and distributing computations in the case that metric params are derived from the dataset (ideally a training set). #16419 is a simple instance where the metric param is sample-aligned and needs to be chunked up. In other cases, we precompute some metric param over all the data, then pass it to each chunk worker, using _precompute_metric_params introduced in #12672. This is also relevant to #9555. While that initial implementation in #12672 is helpful and aims to maintain backwards compatibility, it makes some dubious choices. Firstly in terms of code structure it is not a very modular approach - each metric is handled with an if-then. Secondly, it *only* handles the chunking case, relying on the fact that these metrics are in scipy.spatial, and have a comparable handling of V=None and VI=None. In the Gower Distances PR (#9555) when implementing a metric locally, rather than relying on scipy.spatial, we needed to provide an implementation of these default parameters both when the data is chunked and when the metric function is called straight out. Thirdly, its approach to training vs test data is dubious. We don't formally label X and Y in pairwise_distances as train/test, and perhaps we should. Maintaining backwards compat with scipy's seuclidean and mahalanobis, our implementation stacks X and Y to each other if both are provided, and then calculates their variance. This means that users may be applying a different metric at train and at test time (if the variance of X as train and Y as test is substantially different), which I consider a silent error. We can either make the train/test nature of X and Y more explicit, or we can require that data-based parameters are given explicitly by the user and not implicitly computed. If I understand correctly, sklearn.neighbors will not compute V or VI for you, and it must be provided explicitly. (Requiring that the scaling of each feature be given explicitly in Gower seems like an unnecessary burden on the user, however.) Then there are issues like whether we should consistently set the diagonal to zero in all metrics where Y=None. In short, there are several projects in distances, and I'd support them being considered for work.... But it's a lot of engineering, if motivated by ML needs and consistency for users. J

3 2

Getting identical mse, r2, mae for different data
by Vipula Rawte March 5, 2020

March 5, 2020

I am getting identical metric evaluation values for different data, I printed the matrix shape too. Below is a screenshot: [image: image.png] -- Regards, Vipula Rawte

2 3

ANN: scikit-learn 0.22 final release
by Adrin March 4, 2020

March 4, 2020

We're happy to announce the 0.22 release. You can read the release highlights under https://scikit-learn.org/stable/auto_examples/release_highlights/plot_relea… and the long version of the change log under https://scikit-learn.org/stable/whats_new/v0.22.html#changes-0-22. This version supports Python versions 3.5 to 3.8. You can give it a go using `pip install -U scikit-learn` while conda and conda forge binaries are coming. Regards, Adrin, on behalf of the scikit-learn maintainer team.

5 6

tensorflow and scikit-learn
by Nils Wagner March 3, 2020

March 3, 2020

Hi All, I am newbie to scikit-learn. Is it possible to use scikit-learn instead of tensorflow and keras in the attached script? Best regards, Nils

3 2