Classification predictive modeling problems involve predicting a class label for a given set of inputs.
It is a challenging problem in general, especially if little is known about the dataset, as there are tens, if not hundreds, of machine learning algorithms to choose from.
The problem is made significantly more difficult if the distribution of examples across the classes is imbalanced.
This requires the use of specialized methods to either change the dataset or change the learning algorithm to handle the skewed class distribution.
A common way to deal with the overwhelm on a new classification project is to use a favorite machine learning algorithm like Random Forest or SMOTE.
Another common approach is to scour the research literature for descriptions of vaguely similar problems and attempt to re-implement the algorithms and configurations that are described.
These approaches can be effective, although they are hit-or-miss and time-consuming respectively.
Instead, the shortest path to a good result on a new classification task is to systematically evaluate a suite of machine learning algorithms in order to discover what works well, then double down.
This approach can also be used for imbalanced classification problems, tailored for the range of data sampling, cost-sensitive, and one-class classification algorithms that one may choose from.
In this tutorial, you will discover a systematic framework for working through an imbalanced classification dataset.
After completing this tutorial, you will know:Discover SMOTE, one-class classification, cost-sensitive learning, threshold moving, and much more in my new book, with 30 step-by-step tutorials and full Python source code.
Let’s get started.
Step-By-Step Framework for Imbalanced Classification ProjectsPhoto by ~jar{}, some rights reserved.
This tutorial is divided into three parts; they are:You’re handed or acquire an imbalanced classification dataset.
Now what?There are so many machine learning algorithms to choose from, let alone techniques specifically designed for imbalanced classification.
Which algorithms do you use? How do you choose?This is the challenge faced at the beginning of each new imbalanced classification project.
It is this challenge that makes applied machine learning both thrilling and terrifying.
There are perhaps two common ways to solve this problem:One approach might be to select a favorite algorithm and start tuning the hyperparameters.
This is a fast approach to a solution but is only effective if your favorite algorithm just happens to be the best solution for your specific dataset.
Another approach might be to review the literature and see what techniques have been used on datasets like yours.
This can be effective if many people have studied and reported results on similar datasets.
In practice, this is rarely the case, and research publications are biased towards showing promise for a pet algorithm rather than presenting an honest comparison of methods.
At best, literature can be used for ideas for techniques to try.
Instead, if little is known about the problem, then the shortest path to a “good” result is to systematically test a suite of different algorithms on your dataset.
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-CourseConsider a balanced classification task.
You’re faced with the same challenge of selecting which algorithms to use to address your dataset.
There are many solutions to this problem, but perhaps the most robust is to systematically test a suite of algorithms and use empirical results to choose.
Biases like “my favorite algorithm” or “what has worked in the past” can feed ideas into the study, but can lead you astray if relied upon.
Instead, you need to let the results from systematic empirical experiments to tell you what algorithm is good or best for your imbalanced classification dataset.
Once you have a dataset, the process involves the three steps of (1) selecting a metric by which to evaluate candidate models, (2) testing a suite of algorithms, and (3) tuning the best performing models.
This may not the only approach; it is just the simplest reliable process to get you from “I have a new dataset” to “I have good results” very quickly.
This process can be summarized as follows:Spot-checking algorithms is a little more involved as many algorithms require specialized data preparation such as scaling, removal of outliers, and more.
Also, evaluating candidate algorithms requires the careful design of a test harness, often involving the use of k-fold cross-validation to estimate the performance of a given model on unseen data.
We can use this simple process for imbalanced classification.
It is still important to spot check standard machine learning algorithms on imbalanced classification.
Standard algorithms often do not perform well when the class distribution is imbalanced.
Nevertheless, testing them first provides a baseline in performance by which more specialized models can be compared and must out-perform.
It is also still important to tune the hyperparameters of well-performing algorithms.
This includes the hyperparameters of models specifically designed for imbalanced classification.
Therefore, we can use the same three-step procedure and insert an additional step to evaluate imbalanced classification algorithms.
We can summarize this process as follows:This provides a high-level systematic framework to work through an imbalanced classification problem.
Nevertheless, there are many imbalanced algorithms to choose from, let alone many different standard machine learning algorithms to choose from.
We require a similar low-level systematic framework for each step.
We can develop a similar low-level framework to systematically work through each step of an imbalanced classification project.
From selecting a metric to hyperparameter tuning.
Selecting a metric might be the most important step in the project.
The metric is the measuring stick by which all models are evaluated and compared.
The choice of the wrong metric can mean choosing the wrong algorithm.
That is, a model that solves a different problem from the problem you actually want solved.
The metric must capture those details about a model or its predictions that are most important to the project or project stakeholders.
This is hard, as there are many metrics to choose from and often project stakeholders are not sure what they want.
Three may also be multiple ways to frame the problem, and it may be beneficial to explore a few different framings and, in turn, different metrics to see what makes sense to stakeholders.
First, you must decide whether you want to predict probabilities or crisp class labels.
Recall that for binary imbalanced classification tasks, the majority class is normal, called the “negative class“, and the minority class is the exception, called the “positive class“.
Probabilities capture the uncertainty of the prediction, whereas crisp class labels can be used immediately.
If probabilities are intended to be used directly, then a good metric might be the Brier Score and the Brier Skill score.
Alternately, you may want to predict probabilities and allow the user to map them to crisp class labels themselves via a user-selected threshold.
In this case, a measure can be chosen that summarizes the performance of the model across the range of possible thresholds.
If the positive class is the most important, then the precision-recall curve and area under curve (PR AUC) can be used.
This will optimize both precision and recall across all thresholds.
Alternately, if both classes are equally important, the ROC Curve and area under curve (ROC AUC) can be used.
This will maximize the true positive rate and minimize the false positive rate.
If class labels are required and both classes are equally important, a good default metric is classification accuracy.
This only makes sense if the majority class is less than about 80 percent off the data.
A majority class that has a greater than 80 percent or 90 percent skew will swamp the accuracy metric and it will lose its meaning for comparing algorithms.
If the class distribution is severely skewed, then the G-mean metric can be used that will optimize the sensitivity and specificity metrics.
If the positive class is more important, then variations of the F-Measure can be used that optimize the precision and recall.
If both false positive and false negatives are equally important, then F1 can be used.
If false negatives are more costly, then the F2-Measure can be used, otherwise, if false positives are more costly, then the F0.
5-Measure can be used.
These are just heuristics but provide a useful starting point if you feel lost choosing a metric for your imbalanced classification task.
We can summarize these heuristics into a framework as follows:We can also transform these decisions into a decision tree, as follows.
How to Choose a Metric for Imbalanced ClassificationOnce a metric has been chosen, you can start evaluating machine learning algorithms.
Spot checking machine learning algorithms means evaluating a suite of different types of algorithms with minimal hyperparameter tuning.
Specifically, it means giving each algorithm a good chance to learn about the problem, including performing any required data preparation expected by the algorithm and using best-practice configuration options or defaults.
The objective is to quickly test a range of standard machine learning algorithms and provide a baseline in performance to which techniques specialized for imbalanced classification must be compared and outperform in order to be considered skillful.
The idea here is that there is little point in using fancy imbalanced algorithms if they cannot out-perform so-called unbalanced algorithms.
A robust test harness must be defined.
This often involves k-fold cross-validation, often with k-10 as a sensible default.
Stratified cross-validation is often required to ensure that each fold has the same class distribution as the original dataset.
And the cross-validation procedure is often repeated multiple times, such as 3, 10, or 30 in order to effectively capture a sample of model performance on the dataset, summarized with a mean and standard deviation of the scores.
There are perhaps four levels of algorithms to spot check; they are:Firstly, a naive classification must be evaluated.
This provides a rock-bottom baseline in performance that any algorithm must overcome in order to have skill on the dataset.
Naive means that the algorithm has no logic other than an if-statement or predicting a constant value.
The choice of naive algorithm is based on the choice of performance metric.
For example, a suitable naive algorithm for classification accuracy is to predict the majority class in all cases.
A suitable naive algorithm for the Brier Score when evaluating probabilities is to predict the prior probability of each class in the training dataset.
A suggested mapping of performance metrics to naive algorithms is as follows:If you are unsure of the “best” naive algorithm for your metric, perhaps test a few and discover which results in the better performance that you can use as your rock-bottom baseline.
Some options include:Linear algorithms are those that are often drawn from the field of statistics and make strong assumptions about the functional form of the problem.
We can refer to them as linear because the output is a linear combination of the inputs, or weighted inputs, although this definition is stretched.
You might also refer to these algorithms as probabilistic algorithms as they are often fit under a probabilistic framework.
They are often fast to train and often perform very well.
Examples of linear algorithms you should consider trying include:Nonlinear algorithms are drawn from the field of machine learning and make few assumptions about the functional form of the problem.
We can refer to them as nonlinear because the output is often a nonlinear mapping of inputs to outputs.
They often require more data than linear algorithms and are slower to train.
Examples of nonlinear algorithms you should consider trying include:Ensemble algorithms are also drawn from the field of machine learning and combine the predictions from two or more models.
There are many ensemble algorithms to choose from, but when spot-checking algorithms, it is a good idea to focus on ensembles of decision tree algorithms, given that they are known to perform so well in practice on a wide range of problems.
Examples of ensembles of decision tree algorithms you should consider trying include:We can summarize these suggestions into a framework for testing machine learning algorithms on a dataset.
The order of the steps is probably not flexible.
Think of the order of algorithms as increasing in complexity, and in turn, capability.
The order of algorithms within each step is flexible and the list of algorithms is not complete, and probably never could be given the vast number of algorithms available.
Limiting the algorithms tested to a subset of the most common or most widely used is a good start.
Using data-preparation recommendations and hyperparameter defaults is also a good start.
The figure below summarizes this step of the framework.
How to Spot-Check Machine Learning AlgorithmsSpot-checking imbalanced algorithms is much like spot-checking machine learning algorithms.
The objective is to quickly test a large number of techniques in order to discover what shows promise so that you can focus more attention on it later during hyperparameter tuning.
The spot-checking performed in the previous section provides both naive and modestly skillful models by which all imbalanced techniques can be compared.
This allows you to focus on these methods that truly show promise on the problem, rather than getting excited about results that only appear effective compared only to other imbalanced classification techniques (which is an easy trap to fall into).
There are perhaps four types of imbalanced classification techniques to spot check:Data sampling algorithms change the composition of the training dataset to improve the performance of a standard machine learning algorithm on an imbalanced classification problem.
There are perhaps three main types of data sampling techniques; they are:Data oversampling involves duplicating examples of the minority class or synthesizing new examples from the minority class from existing examples.
Perhaps the most popular methods is SMOTE and variations such as Borderline SMOTE.
Perhaps the most important hyperparameter to tune is the amount of oversampling to perform.
Examples of data oversampling methods include:Undersampling involves deleting examples from the majority class, such as randomly or using an algorithm to carefully choose which examples to delete.
Popular editing algorithms include the edited nearest neighbors and Tomek links.
Examples of data undersampling methods include:Almost any oversampling method can be combined with almost any undersampling technique.
Therefore, it may be beneficial to test a suite of different combinations of oversampling and undersampling techniques.
Examples of popular combinations of over and undersampling include:Data sampling algorithms may perform differently depending on the choice of machine learning algorithm.
As such, it may be beneficial to test a suite of standard machine learning algorithms, such as all or a subset of those algorithms used when spot checking in the previous section.
Additionally, most data sampling algorithms make use of the k-nearest neighbor algorithm internally.
This algorithm is very sensitive to the data types and scale of input variables.
As such, it may be important to at least normalize input variables that have differing scales prior to testing the methods, and perhaps using specialized methods if some input variables are categorical instead of numerical.
Cost-sensitive algorithms are modified versions of machine learning algorithms designed to take the differing costs of misclassification into account when fitting the model on the training dataset.
These algorithms can be effective when used on imbalanced classification, where the cost of misclassification is configured to be inversely proportional to the distribution of examples in the training dataset.
There are many cost-sensitive algorithms to choose from, although it might be practical to test a range of cost-sensitive versions of linear, nonlinear, and ensemble algorithms.
Some examples of machine learning algorithms that can be configured using cost-sensitive training include:Algorithms used for outlier detection and anomaly detection can be used for classification tasks.
Although unusual, when used in this way, they are often referred to as one-class classification algorithms.
In some cases, one-class classification algorithms can be very effective, such as when there is a severe class imbalance with very few examples of the positive class.
Examples of one-class classification algorithms to try include:Predicted probabilities can be improved in two ways; they are:Some algorithms are fit using a probabilistic framework and, in turn, have calibrated probabilities.
This means that when 100 examples are predicted to have the positive class label with a probability of 80 percent, then the algorithm will predict the correct class label 80 percent of the time.
Calibrated probabilities are required from a model to be considered skillful on a binary classification task when probabilities are either required as the output or used to evaluate the model (e.
g.
ROC AUC or PR AUC).
Some examples of machine learning algorithms that predict calibrated probabilities are as follows:Most nonlinear algorithms do not predict calibrated probabilities, therefore algorithms can be used to post-process the predicted probabilities in order to calibrate them.
Therefore, when probabilities are required directly or are used to evaluate a model, and nonlinear algorithms are being used, it is important to calibrate the predicted probabilities.
Some examples of probability calibration algorithms to try include:Some algorithms are designed to naively predict probabilities that later must be mapped to crisp class labels.
This is the case if class labels are required as output for the problem, or the model is evaluated using class labels.
Examples of probabilistic machine learning algorithms that predict a probability by default include:Probabilities are mapped to class labels using a threshold probability value.
All probabilities below the threshold are mapped to class 0, and all probabilities equal-to or above the threshold are mapped to class 1.
The default threshold is 0.
5, although different thresholds can be used that will dramatically impact the class labels and, in turn, the performance of a machine learning model that natively predicts probabilities.
As such, if probabilistic algorithms are used that natively predict a probability and class labels are required as output or used to evaluate models, it is a good idea to try tuning the classification threshold.
We can summarize these suggestions into a framework for testing imbalanced machine learning algorithms on a dataset.
The order of the steps is flexible, and the order of algorithms within each step is also flexible, and the list of algorithms is not complete.
The structure is designed to get you thinking systematically about what algorithm to evaluate.
The figure below summarizes the framework.
How to Spot-Check Imbalanced Machine Learning AlgorithmsAfter spot-checking machine learning algorithms and imbalanced algorithms, you will have some idea of what works and what does not on your specific dataset.
The simplest approach to hyperparameter tuning is to select the top five or 10 algorithms or algorithm combinations that performed well and tune the hyperparameters for each.
There are three popular hyperparameter tuning algorithms that you may choose from:A good default is grid search if you know what hyperparameter values to try, otherwise, random search should be used.
Bayesian optimization should be used if possible but can be more challenging to set up and run.
Tuning the best performing methods is a good start, but not the only approach.
There may be some standard machine learning algorithms that perform well, but do not perform as well when used with data sampling or probability calibration.
These algorithms could be tuned in concert with their imbalanced-classification augmentations to see if better performance can be achieved.
Additionally, there may be imbalanced-classification algorithms, such as a data sampling method that results in a dramatic lift in performance for one or more algorithms.
These algorithms themselves may provide an interesting basis for further tuning to see if additional lifts in performance can be achieved.
This section provides more resources on the topic if you are looking to go deeper.
In this tutorial, you discovered a systematic framework for working through an imbalanced classification dataset.
Specifically, you learned:Do you have any questions? Ask your questions in the comments below and I will do my best to answer.
with just a few lines of python codeDiscover how in my new Ebook: Imbalanced Classification with PythonIt 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.
.