Samplers / pipelines for imbalanced datasets

[TomAugspurger]

Imbalanced datasets, where the classes have very different occurrence rates, can show up in large data sets.

There are many strategies for dealing with imbalanced data. http://contrib.scikit-learn.org/imbalanced-learn/stable/api.html implements a set, some of which could be scaled to large datasets with dask.

[sephib]

Hi, I think that most of the changes would be to introduce the option of fit_resample and fit_sample into the fit_transform method.
I'll be happy to assist on this issue.

[TomAugspurger]

@sephib do you have any examples of fit_resample and fit_sample? I'm not familiar with them.

[sephib]

The core fit_resample function is from within imblearn/base.py.
It is incorporated throughout the imblearn library - for example here is the implementation within imblearn pipeline

[TomAugspurger]

Thanks. The standard sklearn.pipeline.Pipeline works well with dask containers. Does the one in imblearn work with Dask objects? If not, what breaks?

On Tue, Mar 3, 2020 at 3:43 AM sephib notifications@github.com wrote:

The core fit_resample https://github.com/scikit-learn-contrib/imbalanced-learn/blob/6b3c5ae/imblearn/base.py#L54 function is from within imblearn/base.py https://github.com/scikit-learn-contrib/imbalanced-learn/blob/6b3c5ae/imblearn/base.py . It is incorporated throughout the imblearn library - for example here is the implementation within imblearn pipeline https://github.com/scikit-learn-contrib/imbalanced-learn/blob/6b3c5aed61f2e5dc0e8af87d97ea92b95dcafdd0/imblearn/pipeline.py#L333

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/dask/dask-ml/issues/317?email_source=notifications&email_token=AAKAOIXTMCM6CANWCZF4BSLRFTGKRA5CNFSM4FMLFMH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOENSY6EI#issuecomment-593858321, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKAOIUBOVBXDONJBYWE3CLRFTGKRANCNFSM4FMLFMHQ .

[sephib]

Currently when I ran daskml with an imblearn pipeline I got an error:

AttributeError: 'FunctionSamplerw object has no attribute 'transform'

This is from the dask_ml/model_selection/method.py fit_transform function which is looking for a fit_transform or fit and transform attributes (which are in imblearn "converted" to fit_resample

[TomAugspurger]

It'd would help to have a minimal minimal example: http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports

[sephib]

Hi Here is a sample code that passes the dask_ml/model_selection/methods.py . Unfortunately it still does not pass the /imblearn/base.py file but I think it may be something with the example

when amending the file with

from imblearn.pipeline import Pipeline

instead of

from sklearn.pipeline import Pipeline

and adding these lines into the fit_transform function after line 260

elif hasattr(est, "fit_resample"):
                Xt = est.fit_resample(X, y, **fit_params)

rom sklearn.model_selection import train_test_split as tts
from sklearn.datasets import make_classification
from sklearn.decomposition import PCA
from sklearn.metrics import classification_report
from sklearn.neighbors import KNeighborsClassifier as KNN
from imblearn.pipeline import make_pipeline
from imblearn.under_sampling import (EditedNearestNeighbours,
                                     RepeatedEditedNearestNeighbours)
import dask_ml.model_selection as dcv
from sklearn.model_selection import GridSearchCV

# Generate the dataset
X, y = make_classification(n_classes=2, class_sep=1.25, weights=[0.3, 0.7],
                           n_informative=3, n_redundant=1, flip_y=0,
                           n_features=5, n_clusters_per_class=1,
                           n_samples=5000, random_state=10)

# Instanciate a PCA object for the sake of easy visualisation
pca = PCA(n_components=2)

# Create the samplers
enn = EditedNearestNeighbours()
renn = RepeatedEditedNearestNeighbours()

# Create the classifier
knn = KNN(1)

# Make the splits
X_train, X_test, y_train, y_test = tts(X, y, random_state=42)

# Add one transformers and two samplers in the pipeline object
pipeline = make_pipeline(pca, enn, renn, knn)
param_grid = {"pca__n_components":[1, 2, 3],}

# grid = GridSearchCV(pipeline, param_grid=param_grid)
grid = dcv.GridSearchCV(pipeline, param_grid=param_grid)

grid.fit(X_train, y_train)

Any inputs would be appreciated

[TomAugspurger]

Thanks. So the issue is with dask_ml.model_selection.GridSearchCV? I'm confused about how this would work with scikit-learn, since (AFAIK) fit_resample isn't part of their API.

[sephib]

That's the magic of imblearn.pipeline (if you un-comment the dvc.GirdSearchCV and leave the sklearn GridSearchCV the code runs without any errors).

[TomAugspurger]

I don't really see how that would work. But feel free to propose changes in a PR and we can discuss that there.

On Mon, Mar 9, 2020 at 4:50 PM sephib notifications@github.com wrote:

That's the magic of imblearn.pipeline (if you un-comment the dvc.GirdSearchCV and leave the sklearn GridSearchCV the code runs without any errors).

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/dask/dask-ml/issues/317?email_source=notifications&email_token=AAKAOIS7W5XODNWP5WGBPMDRGVQAZA5CNFSM4FMLFMH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOJGJKA#issuecomment-596796584, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKAOIQ7ML7HENASN5PPL5LRGVQAZANCNFSM4FMLFMHQ .

[glemaitre]

@TomAugspurger

I started a POC to adapt our RandomUnderSampler to support natively dask array and dataframe (in/out): https://github.com/scikit-learn-contrib/imbalanced-learn/pull/777

I think that we can do something similar for both RandomOverSampler and ClusterCentroids. They don't rely on kNN and thus make it possible to work in a distributed setting. The other methods rely on kNN and I am not sure that it would be easy to do anything then.

Regarding the integration with the imbalanced-learn Pipeline, our implementation is exactly the one of scikit-learn but we check if a sampler is within the pipeline. This check looks for the attribute fit_resample which would be applied only during fit of the pipeline. Thus, I would say that you can safely use imblearn.Pipeline in replacement of the sklearn.Pipeline.

I was wondering if you would have a bit of time just to check if, on the dask part, we don't implement something stupid in the above PR (I am not super familiar yet with distributed computation).

[sephib]

Regarding the integration with the imbalanced-learn Pipeline, our implementation is exactly the one of scikit-learn but we check if a sampler is within the pipeline. This check looks for the attribute fit_resample which would be applied only during fit of the pipeline. Thus, I would say that you can safely use imblearn.Pipeline in replacement of the sklearn.Pipeline.

@TomAugspurger is a PR still relevant? if so i'll be happy to get some guidance

[TomAugspurger]

I'm not sure what's required, but perhaps imbalanced-learn's Pipeline will just be able to accept Dask collections after that pull request? I don't know what estimators like GridSearchCV need to do (if anything) to work with imbalanced-learn pipelines.

On Fri, Nov 6, 2020 at 7:37 AM sephib notifications@github.com wrote:

Regarding the integration with the imbalanced-learn Pipeline, our implementation is exactly the one of scikit-learn but we check if a sampler is within the pipeline. This check looks for the attribute fit_resample which would be applied only during fit of the pipeline. Thus, I would say that you can safely use imblearn.Pipeline in replacement of the sklearn.Pipeline.

@TomAugspurger https://github.com/TomAugspurger is a PR still relevant? if so i'll be happy to get some guidance

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/dask/dask-ml/issues/317#issuecomment-723084512, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKAOIWQIZFLMFA25KATC5DSOP3YRANCNFSM4FMLFMHQ .

[sephib]

I guess we can see how the @glemaitre PR goes through and then see if there is anything else to do on dask-ml side

[vishalvvs]

Does imblearn supports Dask Natively?? I have been using joblib with parallel_backend = "dask" for it, but it seems that it is not able to parallelize my tasks.

[Jose-Bastos]

Any updates on this? For example, could I use RandomOverSampler if I use @glemaitre 's PR with minor changes? Thank you in advance!