SMOTE for Imbalanced Classification with Python

Last Updated on August 21, 2020

Imbalanced classification involves developing predictive models on classification datasets that have a severe class imbalance.

The challenge of working with imbalanced datasets is that most machine learning techniques will ignore, and in turn have poor performance on, the minority class, although typically it is performance on the minority class that is most important.

One approach to addressing imbalanced datasets is to oversample the minority class. The simplest approach involves duplicating examples in the minority class, although these examples don’t add any new information to the model. Instead, new examples can be synthesized from the existing examples. This is a type of data augmentation for the minority class and is referred to as the Synthetic Minority Oversampling Technique, or SMOTE for short.

In this tutorial, you will discover the SMOTE for oversampling imbalanced classification datasets.

After completing this tutorial, you will know:

  • How the SMOTE synthesizes new examples for the minority class.
  • How to correctly fit and evaluate machine learning models on SMOTE-transformed training datasets.
  • How to use extensions of the SMOTE that generate synthetic examples along the class decision boundary.

Kick-start your project with my new book Imbalanced Classification with Python, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

SMOTE Oversampling for Imbalanced Classification with Python

SMOTE Oversampling for Imbalanced Classification with Python
Photo by Victor U, some rights reserved.

Tutorial Overview

This tutorial is divided into five parts; they are:

  1. Synthetic Minority Oversampling Technique
  2. Imbalanced-Learn Library
  3. SMOTE for Balancing Data
  4. SMOTE for Classification
  5. SMOTE With Selective Synthetic Sample Generation
    1. Borderline-SMOTE
    2. Borderline-SMOTE SVM
    3. Adaptive Synthetic Sampling (ADASYN)

Synthetic Minority Oversampling Technique

A problem with imbalanced classification is that there are too few examples of the minority class for a model to effectively learn the decision boundary.

One way to solve this problem is to oversample the examples in the minority class. This can be achieved by simply duplicating examples from the minority class in the training dataset prior to fitting a model. This can balance the class distribution but does not provide any additional information to the model.

An improvement on duplicating examples from the minority class is to synthesize new examples from the minority class. This is a type of data augmentation for tabular data and can be very effective.

Perhaps the most widely used approach to synthesizing new examples is called the Synthetic Minority Oversampling TEchnique, or SMOTE for short. This technique was described by Nitesh Chawla, et al. in their 2002 paper named for the technique titled “SMOTE: Synthetic Minority Over-sampling Technique.”

SMOTE works by selecting examples that are close in the feature space, drawing a line between the examples in the feature space and drawing a new sample at a point along that line.

Specifically, a random example from the minority class is first chosen. Then k of the nearest neighbors for that example are found (typically k=5). A randomly selected neighbor is chosen and a synthetic example is created at a randomly selected point between the two examples in feature space.

… SMOTE first selects a minority class instance a at random and finds its k nearest minority class neighbors. The synthetic instance is then created by choosing one of the k nearest neighbors b at random and connecting a and b to form a line segment in the feature space. The synthetic instances are generated as a convex combination of the two chosen instances a and b.

— Page 47, Imbalanced Learning: Foundations, Algorithms, and Applications, 2013.

This procedure can be used to create as many synthetic examples for the minority class as are required. As described in the paper, it suggests first using random undersampling to trim the number of examples in the majority class, then use SMOTE to oversample the minority class to balance the class distribution.

The combination of SMOTE and under-sampling performs better than plain under-sampling.

SMOTE: Synthetic Minority Over-sampling Technique, 2011.

The approach is effective because new synthetic examples from the minority class are created that are plausible, that is, are relatively close in feature space to existing examples from the minority class.

Our method of synthetic over-sampling works to cause the classifier to build larger decision regions that contain nearby minority class points.

SMOTE: Synthetic Minority Over-sampling Technique, 2011.

A general downside of the approach is that synthetic examples are created without considering the majority class, possibly resulting in ambiguous examples if there is a strong overlap for the classes.

Now that we are familiar with the technique, let’s look at a worked example for an imbalanced classification problem.

Imbalanced-Learn Library

In these examples, we will use the implementations provided by the imbalanced-learn Python library, which can be installed via pip as follows:

You can confirm that the installation was successful by printing the version of the installed library:

Running the example will print the version number of the installed library; for example:

Want to Get Started With Imbalance Classification?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

Download Your FREE Mini-Course

SMOTE for Balancing Data

In this section, we will develop an intuition for the SMOTE by applying it to an imbalanced binary classification problem.

First, we can use the make_classification() scikit-learn function to create a synthetic binary classification dataset with 10,000 examples and a 1:100 class distribution.

We can use the Counter object to summarize the number of examples in each class to confirm the dataset was created correctly.

Finally, we can create a scatter plot of the dataset and color the examples for each class a different color to clearly see the spatial nature of the class imbalance.

Tying this all together, the complete example of generating and plotting a synthetic binary classification problem is listed below.

Running the example first summarizes the class distribution, confirms the 1:100 ratio, in this case with about 9,900 examples in the majority class and 100 in the minority class.

A scatter plot of the dataset is created showing the large mass of points that belong to the majority class (blue) and a small number of points spread out for the minority class (orange). We can see some measure of overlap between the two classes.

Scatter Plot of Imbalanced Binary Classification Problem

Scatter Plot of Imbalanced Binary Classification Problem

Next, we can oversample the minority class using SMOTE and plot the transformed dataset.

We can use the SMOTE implementation provided by the imbalanced-learn Python library in the SMOTE class.

The SMOTE class acts like a data transform object from scikit-learn in that it must be defined and configured, fit on a dataset, then applied to create a new transformed version of the dataset.

For example, we can define a SMOTE instance with default parameters that will balance the minority class and then fit and apply it in one step to create a transformed version of our dataset.

Once transformed, we can summarize the class distribution of the new transformed dataset, which would expect to now be balanced through the creation of many new synthetic examples in the minority class.

A scatter plot of the transformed dataset can also be created and we would expect to see many more examples for the minority class on lines between the original examples in the minority class.

Tying this together, the complete examples of applying SMOTE to the synthetic dataset and then summarizing and plotting the transformed result is listed below.

Running the example first creates the dataset and summarizes the class distribution, showing the 1:100 ratio.

Then the dataset is transformed using the SMOTE and the new class distribution is summarized, showing a balanced distribution now with 9,900 examples in the minority class.

Finally, a scatter plot of the transformed dataset is created.

It shows many more examples in the minority class created along the lines between the original examples in the minority class.

Scatter Plot of Imbalanced Binary Classification Problem Transformed by SMOTE

Scatter Plot of Imbalanced Binary Classification Problem Transformed by SMOTE

The original paper on SMOTE suggested combining SMOTE with random undersampling of the majority class.

The imbalanced-learn library supports random undersampling via the RandomUnderSampler class.

We can update the example to first oversample the minority class to have 10 percent the number of examples of the majority class (e.g. about 1,000), then use random undersampling to reduce the number of examples in the majority class to have 50 percent more than the minority class (e.g. about 2,000).

To implement this, we can specify the desired ratios as arguments to the SMOTE and RandomUnderSampler classes; for example:

We can then chain these two transforms together into a Pipeline.

The Pipeline can then be applied to a dataset, performing each transformation in turn and returning a final dataset with the accumulation of the transform applied to it, in this case oversampling followed by undersampling.

The pipeline can then be fit and applied to our dataset just like a single transform:

We can then summarize and plot the resulting dataset.

We would expect some SMOTE oversampling of the minority class, although not as much as before where the dataset was balanced. We also expect fewer examples in the majority class via random undersampling.

Tying this all together, the complete example is listed below.

Running the example first creates the dataset and summarizes the class distribution.

Next, the dataset is transformed, first by oversampling the minority class, then undersampling the majority class. The final class distribution after this sequence of transforms matches our expectations with a 1:2 ratio or about 2,000 examples in the majority class and about 1,000 examples in the minority class.

Finally, a scatter plot of the transformed dataset is created, showing the oversampled majority class and the undersampled majority class.

Scatter Plot of Imbalanced Dataset Transformed by SMOTE and Random Undersampling

Scatter Plot of Imbalanced Dataset Transformed by SMOTE and Random Undersampling

Now that we are familiar with transforming imbalanced datasets, let’s look at using SMOTE when fitting and evaluating classification models.

SMOTE for Classification

In this section, we will look at how we can use SMOTE as a data preparation method when fitting and evaluating machine learning algorithms in scikit-learn.

First, we use our binary classification dataset from the previous section then fit and evaluate a decision tree algorithm.

The algorithm is defined with any required hyperparameters (we will use the defaults), then we will use repeated stratified k-fold cross-validation to evaluate the model. We will use three repeats of 10-fold cross-validation, meaning that 10-fold cross-validation is applied three times fitting and evaluating 30 models on the dataset.

The dataset is stratified, meaning that each fold of the cross-validation split will have the same class distribution as the original dataset, in this case, a 1:100 ratio. We will evaluate the model using the ROC area under curve (AUC) metric. This can be optimistic for severely imbalanced datasets but will still show a relative change with better performing models.

Once fit, we can calculate and report the mean of the scores across the folds and repeats.

We would not expect a decision tree fit on the raw imbalanced dataset to perform very well.

Tying this together, the complete example is listed below.

Running the example evaluates the model and reports the mean ROC AUC.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see that a ROC AUC of about 0.76 is reported.

Now, we can try the same model and the same evaluation method, although use a SMOTE transformed version of the dataset.

The correct application of oversampling during k-fold cross-validation is to apply the method to the training dataset only, then evaluate the model on the stratified but non-transformed test set.

This can be achieved by defining a Pipeline that first transforms the training dataset with SMOTE then fits the model.

This pipeline can then be evaluated using repeated k-fold cross-validation.

Tying this together, the complete example of evaluating a decision tree with SMOTE oversampling on the training dataset is listed below.

Running the example evaluates the model and reports the mean ROC AUC score across the multiple folds and repeats.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see a modest improvement in performance from a ROC AUC of about 0.76 to about 0.80.

As mentioned in the paper, it is believed that SMOTE performs better when combined with undersampling of the majority class, such as random undersampling.

We can achieve this by simply adding a RandomUnderSampler step to the Pipeline.

As in the previous section, we will first oversample the minority class with SMOTE to about a 1:10 ratio, then undersample the majority class to achieve about a 1:2 ratio.

Tying this together, the complete example is listed below.

Running the example evaluates the model with the pipeline of SMOTE oversampling and random undersampling on the training dataset.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see that the reported ROC AUC shows an additional lift to about 0.83.

You could explore testing different ratios of the minority class and majority class (e.g. changing the sampling_strategy argument) to see if a further lift in performance is possible.

Another area to explore would be to test different values of the k-nearest neighbors selected in the SMOTE procedure when each new synthetic example is created. The default is k=5, although larger or smaller values will influence the types of examples created, and in turn, may impact the performance of the model.

For example, we could grid search a range of values of k, such as values from 1 to 7, and evaluate the pipeline for each value.

The complete example is listed below.

Running the example will perform SMOTE oversampling with different k values for the KNN used in the procedure, followed by random undersampling and fitting a decision tree on the resulting training dataset.

The mean ROC AUC is reported for each configuration.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, the results suggest that a k=3 might be good with a ROC AUC of about 0.84, and k=7 might also be good with a ROC AUC of about 0.85.

This highlights that both the amount of oversampling and undersampling performed (sampling_strategy argument) and the number of examples selected from which a partner is chosen to create a synthetic example (k_neighbors) may be important parameters to select and tune for your dataset.

Now that we are familiar with how to use SMOTE when fitting and evaluating classification models, let’s look at some extensions of the SMOTE procedure.

SMOTE With Selective Synthetic Sample Generation

We can be selective about the examples in the minority class that are oversampled using SMOTE.

In this section, we will review some extensions to SMOTE that are more selective regarding the examples from the minority class that provide the basis for generating new synthetic examples.

Borderline-SMOTE

A popular extension to SMOTE involves selecting those instances of the minority class that are misclassified, such as with a k-nearest neighbor classification model.

We can then oversample just those difficult instances, providing more resolution only where it may be required.

The examples on the borderline and the ones nearby […] are more apt to be misclassified than the ones far from the borderline, and thus more important for classification.

Borderline-SMOTE: A New Over-Sampling Method in Imbalanced Data Sets Learning, 2005.

These examples that are misclassified are likely ambiguous and in a region of the edge or border of decision boundary where class membership may overlap. As such, this modified to SMOTE is called Borderline-SMOTE and was proposed by Hui Han, et al. in their 2005 paper titled “Borderline-SMOTE: A New Over-Sampling Method in Imbalanced Data Sets Learning.”

The authors also describe a version of the method that also oversampled the majority class for those examples that cause a misclassification of borderline instances in the minority class. This is referred to as Borderline-SMOTE1, whereas the oversampling of just the borderline cases in minority class is referred to as Borderline-SMOTE2.

Borderline-SMOTE2 not only generates synthetic examples from each example in DANGER and its positive nearest neighbors in P, but also does that from its nearest negative neighbor in N.

Borderline-SMOTE: A New Over-Sampling Method in Imbalanced Data Sets Learning, 2005.

We can implement Borderline-SMOTE1 using the BorderlineSMOTE class from imbalanced-learn.

We can demonstrate the technique on the synthetic binary classification problem used in the previous sections.

Instead of generating new synthetic examples for the minority class blindly, we would expect the Borderline-SMOTE method to only create synthetic examples along the decision boundary between the two classes.

The complete example of using Borderline-SMOTE to oversample binary classification datasets is listed below.

Running the example first creates the dataset and summarizes the initial class distribution, showing a 1:100 relationship.

The Borderline-SMOTE is applied to balance the class distribution, which is confirmed with the printed class summary.

Finally, a scatter plot of the transformed dataset is created. The plot clearly shows the effect of the selective approach to oversampling. Examples along the decision boundary of the minority class are oversampled intently (orange).

The plot shows that those examples far from the decision boundary are not oversampled. This includes both examples that are easier to classify (those orange points toward the top left of the plot) and those that are overwhelmingly difficult to classify given the strong class overlap (those orange points toward the bottom right of the plot).

Scatter Plot of Imbalanced Dataset With Borderline-SMOTE Oversampling

Scatter Plot of Imbalanced Dataset With Borderline-SMOTE Oversampling

Borderline-SMOTE SVM

Hien Nguyen, et al. suggest using an alternative of Borderline-SMOTE where an SVM algorithm is used instead of a KNN to identify misclassified examples on the decision boundary.

Their approach is summarized in the 2009 paper titled “Borderline Over-sampling For Imbalanced Data Classification.” An SVM is used to locate the decision boundary defined by the support vectors and examples in the minority class that close to the support vectors become the focus for generating synthetic examples.

… the borderline area is approximated by the support vectors obtained after training a standard SVMs classifier on the original training set. New instances will be randomly created along the lines joining each minority class support vector with a number of its nearest neighbors using the interpolation

Borderline Over-sampling For Imbalanced Data Classification, 2009.

In addition to using an SVM, the technique attempts to select regions where there are fewer examples of the minority class and tries to extrapolate towards the class boundary.

If majority class instances count for less than a half of its nearest neighbors, new instances will be created with extrapolation to expand minority class area toward the majority class.

Borderline Over-sampling For Imbalanced Data Classification, 2009.

This variation can be implemented via the SVMSMOTE class from the imbalanced-learn library.

The example below demonstrates this alternative approach to Borderline SMOTE on the same imbalanced dataset.

Running the example first summarizes the raw class distribution, then the balanced class distribution after applying Borderline-SMOTE with an SVM model.

A scatter plot of the dataset is created showing the directed oversampling along the decision boundary with the majority class.

We can also see that unlike Borderline-SMOTE, more examples are synthesized away from the region of class overlap, such as toward the top left of the plot.

Scatter Plot of Imbalanced Dataset With Borderline-SMOTE Oversampling With SVM

Scatter Plot of Imbalanced Dataset With Borderline-SMOTE Oversampling With SVM

Adaptive Synthetic Sampling (ADASYN)

Another approach involves generating synthetic samples inversely proportional to the density of the examples in the minority class.

That is, generate more synthetic examples in regions of the feature space where the density of minority examples is low, and fewer or none where the density is high.

This modification to SMOTE is referred to as the Adaptive Synthetic Sampling Method, or ADASYN, and was proposed to Haibo He, et al. in their 2008 paper named for the method titled “ADASYN: Adaptive Synthetic Sampling Approach For Imbalanced Learning.”

ADASYN is based on the idea of adaptively generating minority data samples according to their distributions: more synthetic data is generated for minority class samples that are harder to learn compared to those minority samples that are easier to learn.

ADASYN: Adaptive synthetic sampling approach for imbalanced learning, 2008.

With online Borderline-SMOTE, a discriminative model is not created. Instead, examples in the minority class are weighted according to their density, then those examples with the lowest density are the focus for the SMOTE synthetic example generation process.

The key idea of ADASYN algorithm is to use a density distribution as a criterion to automatically decide the number of synthetic samples that need to be generated for each minority data example.

ADASYN: Adaptive synthetic sampling approach for imbalanced learning, 2008.

We can implement this procedure using the ADASYN class in the imbalanced-learn library.

The example below demonstrates this alternative approach to oversampling on the imbalanced binary classification dataset.

Running the example first creates the dataset and summarizes the initial class distribution, then the updated class distribution after oversampling was performed.

A scatter plot of the transformed dataset is created. Like Borderline-SMOTE, we can see that synthetic sample generation is focused around the decision boundary as this region has the lowest density.

Unlike Borderline-SMOTE, we can see that the examples that have the most class overlap have the most focus. On problems where these low density examples might be outliers, the ADASYN approach may put too much attention on these areas of the feature space, which may result in worse model performance.

It may help to remove outliers prior to applying the oversampling procedure, and this might be a helpful heuristic to use more generally.

Scatter Plot of Imbalanced Dataset With Adaptive Synthetic Sampling (ADASYN)

Scatter Plot of Imbalanced Dataset With Adaptive Synthetic Sampling (ADASYN)

Further Reading

This section provides more resources on the topic if you are looking to go deeper.

Books

Papers

API

Articles

Summary

In this tutorial, you discovered the SMOTE for oversampling imbalanced classification datasets.

Specifically, you learned:

  • How the SMOTE synthesizes new examples for the minority class.
  • How to correctly fit and evaluate machine learning models on SMOTE-transformed training datasets.
  • How to use extensions of the SMOTE that generate synthetic examples along the class decision boundary.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Get a Handle on Imbalanced Classification!

Imbalanced Classification with Python

Develop Imbalanced Learning Models in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Imbalanced Classification with Python

It provides self-study tutorials and end-to-end projects on:
Performance Metrics, Undersampling Methods, SMOTE, Threshold Moving, Probability Calibration, Cost-Sensitive Algorithms
and much more...

Bring Imbalanced Classification Methods to Your Machine Learning Projects

See What's Inside

135 Responses to SMOTE for Imbalanced Classification with Python

  1. Markus January 17, 2020 at 10:52 pm #

    Hi

    print(‘Mean ROC AUC: %.3f’ % mean(scores))

    For calculatng ROC AUC, the examples make use of the mean function an not roc_auc_score, why?

    Thanks

    • Jason Brownlee January 18, 2020 at 8:48 am #

      The ROC AUC scores are calculated automatically via the cross-validation process in scikit-learn.

      • Ram pratapa April 1, 2020 at 6:13 pm #

        Hi Jason,
        Is there any way to use smote for multilabel problem.

        • Jason Brownlee April 2, 2020 at 5:44 am #

          Yes, you must specify to the smote config which are the positive/negative clasess and how much to oversample them.

  2. Camara Mamadou January 21, 2020 at 12:52 am #

    Hi Jason,

    thanks for sharing machine learning knowledge.

    How to get predictions on a holdout data test after getting best results of a classifier by SMOTE oversampling?

    Best regards!

    Mamadou.

    • Jason Brownlee January 21, 2020 at 7:15 am #

      Call model.predict() as per normal.

      Recall SMOTE is only applied to the training set when your model is fit.

      • Akil February 20, 2020 at 11:47 pm #

        Hi Jason,

        As you said, SMOTE is applied to training only, won’t that affect the accuracy of the test set?

        • Jason Brownlee February 21, 2020 at 8:23 am #

          Yes, the model will have a better idea of the boundary and perform better on the test set – at least on some datasets.

  3. Rafael Eder January 21, 2020 at 3:17 pm #

    Hi !
    SMOTE works for imbalanced image datasets too ?

    Best Regards;

    • Jason Brownlee January 22, 2020 at 6:17 am #

      No, it is designed for tabular data.

      You might be able to use image augmentation in the same manner.

      • Rafael Eder January 22, 2020 at 10:07 am #

        Yours books and blog help me a lot ! Thank you very much !

  4. brian January 31, 2020 at 12:28 am #

    Hi Jason, thanks for another series of excellent tutorials. I have encountered an error when running

    X, y = pipeline.fit_resample(X, y)

    on my own X & y imbalanced data. The error is :

    “ValueError: The specified ratio required to remove samples from the minority class while trying to generate new samples. Please increase the ratio.”

    from _validation.py”, line 362, in _sampling_strategy_float in imblearn/utils in the library.

    Could you or anyone else shed some light on this error?

    Thanks.

    • brian January 31, 2020 at 1:50 am #

      as a followup it seems I’ve not understood how SMOTE and undersampling function.

      My input data size is:

      {0: 23558, 1: 8466}

      so a little under 1:3 for minority:majority examples of the classes

      Now I understand I had the ratios for SMOTE() and RandomUnderSampler() “sampling_strategy” incorrect.

      Onwards and upwards!

      • Jason Brownlee January 31, 2020 at 7:57 am #

        Happy to hear that, nice work!

      • Volkan Yurtseven July 22, 2020 at 7:32 am #

        Hi
        When used with a gridsearchcv, does Smote apply the oversampling to whole train set or does it disregard the validation set?

        • Jason Brownlee July 22, 2020 at 7:38 am #

          You can use it as part of a Pipeline to ensure that SMOTE is only applied to the training dataset, not val or test.

    • Jason Brownlee January 31, 2020 at 7:55 am #

      Confirm you have examples of both classes in the y.

    • Jeong miae March 21, 2020 at 10:56 am #

      Thank you for your tutorial.
      I’d like to ask several things.

      1. Could I apply this sampling techniques to image data?

      2. After making balanced data with these thechniques, Could I use not machine learning algorithms but deep learning algorithms such as CNN?

  5. Valdemar February 11, 2020 at 2:06 am #

    Hello Jason,

    In your ML cheat sheet you have advice to invent more data if you have not enough. Can you suggest methods or libraries which are good fit to do that?

    Imblearn seams to be a good way to balance data. What about if you wish to increase the entire dataset size as to have more samples and potentially improve model?

  6. Frank February 28, 2020 at 11:41 pm #

    Thank you for the great tutorial, as always super detailed and helpfull.

    I’m working throught the wine quality dataset(white) and decided to use SMOTE on Output feature balances are below.

    {6: 2198, 5: 1457, 7: 880, 8: 175, 4: 163, 3: 20, 9: 5}

    In your opinion would it be possible to apply SMOTE in this multiclass problem?

    I’ve managed to use a Regression model (KNN) that I belive does the task well but interested to get your take how to deal with similar class imbalance on multilclass problems as above?

    • Jason Brownlee February 29, 2020 at 7:13 am #

      Yes, SMOTE can be used for multi-class, but you must specify the positive and negative classes.

  7. Thomas March 1, 2020 at 6:33 am #

    Thanks for sharing Jason.
    In imblearn.pipeline the predict method says tahar it applies transforms AND sampling and then the final predict of the estimator.

    Therefore isnt that a problem in crossvalscore the sampling will be applied on each validation sets ?

    Thanks

    • Jason Brownlee March 2, 2020 at 6:07 am #

      Sorry, I don’t follow your question. Can you please rephrase or elaborate?

  8. Yong March 1, 2020 at 6:19 pm #

    you mentioned that : ” As in the previous section, we will first oversample the minority class with SMOTE to about a 1:10 ratio, then undersample the majority class to achieve about a 1:2 ratio.”
    why? what is the idea behind this operation and why does this operation can inprove the performance.

    • Jason Brownlee March 2, 2020 at 6:16 am #

      This approach can be effective. It is important to try a range of approaches on your dataset to see what works best.

  9. Vijay M March 2, 2020 at 8:37 pm #

    Sir Jason,
    Can we use the above code for images

  10. Ernest Montañà March 18, 2020 at 3:40 am #

    Hello Jason, thanks for the tutorial.

    When using the lines:

    # define pipeline
    steps = [(‘over’, SMOTE()), (‘model’, RandomForestClassifier(n_estimators=100, criterion=’gini’, max_depth=None, random_state=1))]
    pipeline = Pipeline(steps=steps)
    # evaluate pipeline
    cv = RepeatedStratifiedKFold(n_splits=5, n_repeats=2, random_state=1)
    acc = cross_val_score(pipeline, X_new, Y, scoring=’accuracy’, cv=cv, n_jobs=-1)

    I assume the SMOTE is performed for each cross validation split, therefore there is no data leaking, am I correct? Thank you

  11. David March 29, 2020 at 1:35 am #

    Hi! A quick question, SMOTE should be applied before or after data preparation (like Standardization for example) ? Or it’s irrelevant?

    Thank you!

    • Jason Brownlee March 29, 2020 at 6:01 am #

      Probably after.

      It is doing a knn, so data should be scaled first.

  12. San April 2, 2020 at 6:13 am #

    How to use SMOTE or any other technique related with SMOTE such as ADASYN, Borderline SMOTE, when a dataset has classes with only a few instances?

    Some of the classes in my dataset has only 1 instance & some have 2 instances. When using these SMOTE techniques I get the error ‘Expected n_neighbors <= n_samples, but n_samples = 2, n_neighbors = 6'.

    Is there any way to overcome this error? With RandomOversampling the code works fine..but it doesn't seem to give a good performance. And I'm unable to all the SMOTE based oversampling techniques due to this error.

    • Jason Brownlee April 2, 2020 at 6:41 am #

      I don’t think modeling a problem with one instance or a few instances of a class is appropriate.

      Perhaps collect more data?
      Perhaps delete the underrepresented classes?
      Perhaps reframe the problem?

  13. Garv April 8, 2020 at 9:59 pm #

    Hello I did tuning of smote parameters( k,sampling strategy) and took roc_auc as scoring on training data but how along with cross val score my model is evaluated on testing data (that ideally should not be the one on which smote should apply)
    can you help me with how to apply best model on testing data(code required)
    #Using Decsion Tree
    Xtrain1=Xtrain.copy()
    ytrain1=ytrain.copy()
    k_val=[i for i in range(2,9)]
    p_proportion=[i for i in np.arange(0.2,0.5,0.1)]
    k_n=[]
    proportion=[]
    score_m=[]
    score_var=[]
    modell=[]
    for k in k_val:
    for p in p_proportion:
    oversample=SMOTE(sampling_strategy=p,k_neighbors=k,random_state=1)
    Xtrain1,ytrain1=oversample.fit_resample(Xtrain,ytrain)
    model=DecisionTreeClassifier()
    cv=RepeatedStratifiedKFold(n_splits=10,n_repeats=3,random_state=1)
    scores=cross_val_score(model,X1,y1,scoring=’roc_auc’,cv=cv,n_jobs=-1)
    k_n.append(k)
    proportion.append(p)
    score_m.append(np.mean(scores))
    score_var.append(np.var(scores))
    modell.append(‘DecisionTreeClassifier’)
    scorer=pd.DataFrame({‘model’:modell,’k’:k_n,’proportion’:proportion,’scores’:score_m,’score_var’:score_var})
    print(scorer)
    models.append(model)
    models_score.append(scorer[scorer[‘scores’]==max(scorer[‘scores’])].values[0])
    models_var.append(scorer[scorer[‘score_var’]==min(scorer[‘score_var’])].values[0])

  14. Kabilan April 10, 2020 at 7:02 am #

    Hey Jason,
    What kind of an approach can we use to over-sample time series data?

    • Jason Brownlee April 10, 2020 at 8:38 am #

      Good question, I hope I can cover that topic in the future.

      • John White April 10, 2020 at 7:11 pm #

        Hello Jason,

        Do you currently have any ideas on how to oversample time series data off the top of your head? I’d like to do some research/experiment on it in the meantime.Thank you!

        • Jason Brownlee April 11, 2020 at 6:14 am #

          No, in general I rather make recommendations after doing my homework.

      • Kabilan April 10, 2020 at 11:40 pm #

        Thank you very much!

  15. Vamshi April 11, 2020 at 7:56 am #

    Hi Jason Brownie,

    Thank you for the great description over handling imbalanced datasets using SMOTE and its alternative methods. I know that SMOTE is only for multi Class Dataset but I am curious to know if you have any idea of of using SMOTE for multi label Datasets?? or Do you have any other method or ideas apart from SMOTE in order to handle imbalanced multi label datasets.

    • Jason Brownlee April 11, 2020 at 7:58 am #

      Great question!

      I’m not aware of an approach off hand for multi-label, perhaps check the literature?

      • Vamshi April 11, 2020 at 8:10 am #

        I was working on a dataset as a part of my master thesis and it is highly imbalanced. So I tried experimenting directly using OnevsRestClassifier(without any oversampling) and naturally the classsifer gave worst results(the target value with high number of occurences is being predicted). So I tried testing with Random forest classifier taking each target column one at a time and oversampled with a randomsampler class which gave decent results after oversampling. And I am not sure if I can do it in this way.

  16. Vamshi April 11, 2020 at 8:21 am #

    I also found this solution.
    https://github.com/scikit-learn-contrib/imbalanced-learn/issues/340

  17. Jooje April 16, 2020 at 4:23 am #

    Hi! Thanks for the great tutorial. Can SMOTE be used with 1. high dimensional embeddings for text representation? if so, what is any preprocessing/dimensionality reduction required before applying SMOTE?

    • Jason Brownlee April 16, 2020 at 6:06 am #

      Not sure off the cuff, perhaps experiment to see if this makes sense.

  18. rahul malik April 23, 2020 at 8:49 am #

    hi Jason , I am having 3 input Text columns out of 2 are categorical and 1 is unstructured text. Can you please help me how to do sampling. Output column is categorical and is imbalanced.

    • Jason Brownlee April 23, 2020 at 1:34 pm #

      Perhaps use a label or one hot encoding for the categorical inputs and a bag of words for the text data.

      You can see many examples on the blog, try searching.

      • rahul malik April 23, 2020 at 10:53 pm #

        I have used Pipeline and columntransformer to pass multiplecolumns as X but for sampling I ma not to find any example.For single column I ma able to use SMOTE but how to pass more than in X?

        • Jason Brownlee April 24, 2020 at 5:43 am #

          You may have to experiment, perhaps different smote instances, perhaps run the pipeline manually, etc.

  19. Iraj April 30, 2020 at 8:28 am #

    Hi,
    SMOTE requires 6 examples of each class.
    I have a dataset if 30 class 0, and 1 class 1 .
    Please advise if any solution.
    Thank you

  20. John Sammut May 2, 2020 at 9:25 am #

    Hello Jason,

    Many thanks for this article. I found it very interesting.

    How can one apply the same ratio of oversampling (1:10) followed by under-sampling (1:2) in a pipeline when there are 3 classes?

    The sampling strategy cannot be set to float for multi-class. What would you recommend?

    Thank you.
    John

    • Jason Brownlee May 3, 2020 at 6:05 am #

      Thanks.

      First step is to group classes into positive and negative, then apply the sampling.

  21. Srisha May 4, 2020 at 12:35 pm #

    Could you shed some light on how one could leverage the parameter sampling_strategy in SMOTE?

  22. Mohamad May 7, 2020 at 11:04 pm #

    Hi Jason,
    Thank you very much for this article, it’s so helpful (as always).

    I have an inquiry:
    Now my data are highly imbalanced (99.5%:0.05%). I am having over than 40,000 samples with multiple features (36) for my classification problem. I oversampled with SMOTE to have balanced data, but the classifier is getting highly biased toward the oversampled data. I assumed that its because of the “sampling_strategy”. So I tried {0.25, 0.5, 0.75,1} for the “sampling_strategy”. Its either getting highly biased towards the abundant or the rare class.

    What do you think could be the problem?

  23. john sen May 11, 2020 at 3:45 am #

    please tell me how i can apply two balancing technique first SMOTE and then one class learning algorithm on same dataset for better result

    • Jason Brownlee May 11, 2020 at 6:08 am #

      You can apply smote to the training set, then apply the one class classifier directly.

      I don’t expect it would be beneficial to combine these two methods.

      • john sen May 12, 2020 at 3:54 am #

        sir then what should i try for the best result by using smote and one more algo which makes an hybrid approch to handle imbalanced data.

        • Jason Brownlee May 12, 2020 at 6:50 am #

          Use trial and error to discover what works well/best for your dataset.

  24. Arnaud May 11, 2020 at 3:47 am #

    Hi Jason,
    First, thanks for your material, it’s of great value!

    I have a supervised classification problem with unbalanced class to predict (Event = 1/100 Non Event).

    I have the intuition that using resampling methods such as SMOTE (or down/up/ROSE) with Naive Bayes models affect prior probabilities and such lead to lower performance when applied on test set.

    Is that correct?
    Thanks.

  25. Teixeira May 12, 2020 at 1:31 am #

    Hi Dr.

    Could SMOTE be applied to data that will be used for feeding an LSTM? (Since the order matters, it can interfere with the data right?)

    Thanks in advance!

    • Jason Brownlee May 12, 2020 at 6:46 am #

      No, it is for tabular data only.

      • Teixeira May 13, 2020 at 2:04 am #

        First of all, thanks for the response. Sorry, i think i don’t understand. Maybe I am wrong, but SMOTE could be applied to tabular data, before the transformation into sliding windows. Even in this case is not recommend to apply SMOTE ?

        Thanks!

        • Jason Brownlee May 13, 2020 at 6:39 am #

          Perhaps, but I suspect data generation methods that are time-series-aware would perform much better.

  26. John D May 14, 2020 at 10:04 am #

    Jason,

    I have a highly imbalanced binary (yes/no) classification dataset. The dataset currently has appx 0.008% ‘yes’.

    I need to balance the dataset using SMOTE.

    I came across 2 method to deal with the imbalance. The following steps after I have run MinMaxScaler on the variables

    from imblearn.pipeline import Pipeline
    oversample = SMOTE(sampling_strategy = 0.1, random_state=42)
    undersample = RandomUnderSampler(sampling_strategy=0.5, random_state=42)
    steps = [(‘o’, oversample), (‘u’, undersample)]
    pipeline = Pipeline(steps=steps)
    x_scaled_s, y_s = pipeline.fit_resample(X_scaled, y)
    This results in a reduction in the size of the dataset from 2.4million rows to 732000 rows And the imbalance improves from 0.008% to 33.33%

    While this approach

    sm = SMOTE(random_state=42)
    X_sm , y_sm = sm.fit_sample(X_scaled, y)
    This increases the number of rows from 2.4million rows to 4.8 million rows and the imbalance is now 50%.

    After these steps I need to split data into Train Test datasets….

    What is the right approach here?

    What factors do I need to consider before I choose any of these methods?

    Should I run the X_test, y_test on unsampled data. This would mean, I split the data and do upsampling/undersampling only on the train data.

    Thanks again.

    • Jason Brownlee May 14, 2020 at 1:27 pm #

      No, the sampling is applied on the training dataset only, not the test set. E.g. split first then sample.

  27. Shivam May 16, 2020 at 4:50 pm #

    Hello Jason, Great article. One Issue i am facing while using SMOTE-NC for categorical data. I have only feature for categorization.

    from imblearn.over_sampling import SMOTE
    from imblearn.over_sampling import SMOTENC

    sm = SMOTENC(random_state=27,categorical_features=[0,])

    X_new = np.array(X_train.values.tolist())
    Y_new = np.array(y_train.values.tolist())

    print(X_new.shape) # (10500,)
    print(Y_new.shape) # (10500,)

    X_new = np.reshape(X_new, (-1, 1)) # SMOTE require 2-D Array, Hence changing the shape of X_mew

    print(X_new.shape) # (10500, 1)

    sm.fit_sample(X_new, Y_new)

    But i am getting Error:

    ValueError: Found array with 0 feature(s) (shape=(10500, 0)) while a minimum of 1 is required.

    Can you please suggest how to deal with SMOTE if there is only one feature ?

    • Jason Brownlee May 17, 2020 at 6:31 am #

      Interesting, I wonder if it is a bug in smote-nc?

      Perhaps try duplicating the column and whether it makes a difference?

  28. sukhpal May 26, 2020 at 7:23 pm #

    sir how SMOTE can be applied on CSV file data

  29. John D May 27, 2020 at 8:19 am #

    What is the criteria to UnderSample the majority class and Upsample the minority class.
    OR
    What is the criteria to Upsample the minority class only.

    • Jason Brownlee May 27, 2020 at 1:25 pm #

      I don’t approach it that way. I think it’s misleading and intractable.

      Instead, I recommend do the experiment and use it if it results in better performance.

  30. SUKHPAL May 28, 2020 at 3:51 pm #

    SIR PLEASE PROVIDE TUTORIAL ON TEST TIME AUGMENTATION FOR NUMERICAL DATA

    • Jason Brownlee May 29, 2020 at 6:20 am #

      No problem, I have one written and scheduled to appear next week.

  31. sukhpal May 30, 2020 at 7:01 pm #

    Sir is we apply feature selection technique first or data augmentation first.

  32. Suyash June 25, 2020 at 10:32 pm #

    Why are we implementing SMOTE on whole dataset “X, y = oversample.fit_resample(X, y)”? We should apply oversampling only on the training set. Am i right? What should be done to implement oversampling only on the training set and we also want to use stratified approach?

    • Jason Brownlee June 26, 2020 at 5:35 am #

      Correct, and we do that later in the tutorial when evaluating models.

      In the first example I am getting you used to the API and the affect of the method.

      • suyash June 27, 2020 at 11:38 pm #

        Can you please refer that tutorial to me where we we are implementing smote on taining data only and evaluating the model? I also want to know that RepeatedStratifiedKfold works only on the training dataset only.

      • suyash June 28, 2020 at 12:10 am #

        cross_val_score oversample the data of training set only and do not oversample the training data. am i right?

        • Jason Brownlee June 28, 2020 at 5:52 am #

          When using a pipeline the transform is only applied to the training dataset, which is correct.

          • suyash June 29, 2020 at 3:44 pm #

            Thank You very much.

          • Jason Brownlee June 30, 2020 at 6:11 am #

            You’re welcome.

  33. Jose Q June 28, 2020 at 7:29 am #

    Hi Jason!
    Thank you for such a great post!

    I am working with an imbalanced data set (500:1). I want to get the best recall performance and I have tried with several classification algorithms, hyper parameters, and Over/Under sampling techniques. I will try SMOTE now !!!
    From the last question, I understand that using CV and pipelines you oversample only the training set, right?

    I have another question. My imbalanced data set is about 5 million records from 11 months. It is not a time series. I used data from the first ten months for training, and data from the eleventh month for testing in order to explain it easier to my users, but I feel that it is not correct, and I guess I should use a random test split from the entire data set. Is this correct?

  34. Jose Q July 1, 2020 at 3:09 am #

    Hi Jason,
    I followed your ideas at:
    https://machinelearningmastery.com/framework-for-imbalanced-classification-projects/

    I tried oversampling with SMOTE, but my computer just can’t handle it.
    Then I tried using Decision Trees and XGB for imbalanced data sets after reading your posts:
    https://machinelearningmastery.com/cost-sensitive-decision-trees-for-imbalanced-classification/
    https://machinelearningmastery.com/xgboost-for-imbalanced-classification/
    but I still get low values for recall.

    I am doing random undersample so I have 1:1 class relationship and my computer can manage it. Then I am doing XGB/Decision trees varying max_depth and varying weight to give more importance to the positive class. My assumption is that I won’t overfit the model as soon as I use CV with several folds and iterations. Is that right?
    Thanks

    • Jason Brownlee July 1, 2020 at 5:55 am #

      Well don on your progress!

      Perhaps. Assumptions can lead to poor results, test everything you can think of.

  35. xplorer4us July 8, 2020 at 5:32 pm #

    Hi Jason, excellent explanations on SMOTE, very easy to understand and with tons of examples!

    I tried to download the free mini-course on Imbalance Classification, and I didn’t receive the PDF file.

    May I please ask for your help with this? Thanks in advance!

  36. Landry July 13, 2020 at 8:01 pm #

    Hi Jason, thanks for this tutorial it’s so useful as usual,

    I have one inquiry, I have intuition that SMOTE performs bad on dataset with high dimensionality i.e when we have many features in our dataset. Is it true ?

    • Jason Brownlee July 14, 2020 at 6:18 am #

      Hmmm, that would be my intuition too, but always test. Intuitions breakdown in high dimensions, or with ml in general. Test everything.

  37. Volkan Yurtseven July 22, 2020 at 7:34 am #

    Hi
    When used with a gridsearchcv, does Smote apply the oversampling to whole train set or does it disregard the validation set?

    • Jason Brownlee July 22, 2020 at 7:38 am #

      You can use it as part of a Pipeline to ensure that SMOTE is only applied to the training dataset, not val or test.

      • Volkan Yurtseven July 23, 2020 at 6:52 am #

        hi jason,
        do you mean if i use it in a imblearn’s own Pipeline class, it would be enough? no need for any parameter?

        pipe = Pipeline(steps=[(‘coltrans’, coltrans),
        (‘scl’,StandardScaler()),
        (‘smote’, SMOTE(random_state=42))
        ]
        )

        X_smote,y_smote=pipe.fit_resample(X_train,y_train)

  38. Diego July 23, 2020 at 12:39 pm #

    Hi Jason,

    Thanks for sharing. It really helps in my work 🙂

    Let’s say you train a pipeline using a train dataset and it has 3 steps: MinMaxScaler, SMOTE and LogisticRegression.

    Can you use the same pipeline to preprocess test data ?
    Or should you have a different pipleine without smote for test data ?

    How does pipeline.predict(X_test) that it should not execute SMOTE ?

    Thanks.

  39. SAM V July 29, 2020 at 3:52 pm #

    Hi Jason, SMOTE sampling is done before / after data cleaning or pre-processing or feature engineering??? I just want to know when we should do SMOTE sampling and why??

    • Jason Brownlee July 30, 2020 at 6:16 am #

      It depends on what data prep you are doing.

      Probably after.

  40. Gaël August 6, 2020 at 7:26 pm #

    Hi, great article! I think there is a typo in section “SMOTE for Balancing Data”: “the large mass of points that belong to the minority class (blue)” –> should be majority I guess

  41. Luna August 6, 2020 at 7:30 pm #

    Hi Jason,

    TypeError: All intermediate steps should be transformers and implement fit and transform or be the string ‘passthrough’ ‘SMOTE(k_neighbors=5, n_jobs=None, random_state=None, sampling_strategy=’auto’)’ (type ) doesn’t

    I get this error when running GridSearchCV. What is wrong?

    • Jason Brownlee August 7, 2020 at 6:24 am #

      Perhaps confirm the content of your pipeline ends with a predictive model.

  42. george August 12, 2020 at 1:30 pm #

    Hi Jason,

    if all my predictors are binary, can I still use SMOTE? seems SMOTE only works for predictors are numeric? Are there any methods other than random undersampling or over sampling? Thanks

  43. Franco August 19, 2020 at 7:18 am #

    Brilliant post Jason!

    I wonder if we upsampled the minority class from 100 to 9,900 with a bootstrap (with replacement of course), whether we would get similar results than SMOTE … I put on my to-do list.

    • Jason Brownlee August 19, 2020 at 1:34 pm #

      Thanks!

      Probably not, as we are generating entirely new samples with SMOTE. Nevertheless, run the experiment and compare the results!

  44. Franco August 19, 2020 at 3:52 pm #

    Interesting …. I will. Thanks Jason!

  45. SaHaR August 22, 2020 at 12:35 am #

    Hi, Jason

    Thank you for your great article. It is really informative as always. Recently I read an article about the classification of a multiclass and imbalanced dataset. They used SMOTE for both training and test set and I think it was not a correct methodology and the test dataset should not be manipulated. please tell me if I am wrong and would you recommend a reference about the drawbacks and challenges of using SMOTE?

    Thank you

    • Jason Brownlee August 22, 2020 at 6:17 am #

      Thanks!

      Agreed, it is invalid to use SMOTE on the test set.

  46. Vivek August 28, 2020 at 4:04 am #

    Hi Jason

    Q1. Do we apply SMOTE on the train set after doing train/ test split?

    Guess, doing SMOTE first, then splitting, may result in data leak as same instances may be present in both test and test sets.

    Q2. I understand why SMOTE is better instead of random oversampling minority class. But say for a class imbalance of 1:100, why not just random undersample majority class? Not sure how SMOTE helps here !

    Thanks
    Vivek

    • Jason Brownlee August 28, 2020 at 6:55 am #

      Yes. Training set only.

      Try many methods and discover what works best for your dataset.

  47. Shehab August 29, 2020 at 7:13 am #

    Hi Jason,

    What if you have an unbalanced dataset that matches the realistic class distribution in production. Say Class A has 1000 rows, Class B 400 and Class C with 60. What are the negative effects of having an unbalanced dataset like this. Say I use a classifier like Naive Bayes and since prior probability is important then by oversampling Class C I mess up the prior probability and stray farther away from the realistic probabilities in production. Should I try and get more data or augment the data that I have while maintaining this unbalanced distribution or change the distribution by oversampling the minority classes?

    Thanks

  48. Daniel September 10, 2020 at 7:34 pm #

    Hello,

    Thanks for your work, it is really useful. I have a question about the combination of SMOTE and active learning.

    I am trying to generate a dataset using active learning techniques. From a pool of unlabelled data I select the new points to label using the uncertainty in each iteration. My problem is that the classes repartition is imbalanced (1000:1), my current algorithm can’t find enough points in Yes class. Do you think I could use SMOTE to generate new points of Yes class?
    I am thinking about using borderline-SMOTE to generate new points and then label them. How can I be sure that the new points are not going to be concentrated in a small region?

    I am not sure if I have explained the problem well. I need to find the feasible zone using the labeller in a smart way because labelling is expensive. Can you give me any advice?

    Thanks.

    Daniel

Leave a Reply