Increasing successful detections using data resampling
1. Increasing successful detections using data resampling
In this video, we're going to talk about data resampling methods. They help us to better train our models to recognize fraud cases when there are only very few cases of fraud.2. Undersampling
The most straightforward way to adjust the imbalance of your data, is to undersample the majority class, aka non-fraud cases, or oversample the minority class, aka the fraud cases. With undersampling, you take random draws from your non-fraud observations, to match the amount of fraud observations as seen on the picture.3. Oversampling
With oversampling, you take random draws from the fraud cases, and copy those observations to increase the amount of fraud samples you have in your data. Both methods lead to having a perfect balance between fraud and non-fraud data. But there are drawbacks. With random undersampling, you are effectively throwing away a lot of data and information. With oversampling, you are simply copying data, and therefore are training your model on a lot of duplicates. Let's see how you can implement these methods in practice.4. Oversampling in Python
You can implement resampling methods using Python's imbalanced learn module. It is compatible with scikit-learn and allows you to implement these methods in just two lines of code. As you can see here, I import the package and take the Random Oversampler and assign it to method. I simply fit the method onto my original feature set X, and labels y, to obtained a resampled feature set X, and resampled y. I plot the datasets here side by side, such that you can see the effect of my resampling method. The darker blue color of the data points reflect that there are more identical data points now.5. Synthetic Minority Oversampling Technique (SMOTE)
The Synthetic Minority Oversampling Technique, or SMOTE, is another way of adjusting the imbalance by oversampling your minority observations, aka your fraud cases. But with SMOTE, we're not just copying the minority class. Instead, as you see in this picture, SMOTE uses characteristics of nearest neighbors of fraud cases to create new synthetic fraud cases, and thereby avoids duplicating observations.6. Which resampling method to use?
You might wonder which one of these methods is the best? Well, it depends very much on the situation. If you have very large amounts of data, and also many fraud cases, you might find it computationally easier to undersample, rather than to increase data even more. But in most cases, throwing away data is not desirable. When it comes to oversampling, SMOTE is more sophisticated as it does not duplicate data. But this only works well if your fraud cases are quite similar to each other. If fraud is spread out over your data and not very distinct, using nearest neighbors to create more fraud cases introduces a bit of noise in the data, as the nearest neighbors might not necessarily be fraud cases.7. When to use resampling methods
One thing to keep in mind when using resampling methods, is to only resample on your training set. Your goal is to better train your model by giving it balanced amounts of data. Your goal is not to predict your synthetic samples. Always make sure your test data is free of duplicates or synthetic data, such that you can test your model on real data only. The way to do this, is to first split the data into a train and test set, as you can see here. I then resample the training set only. I fit my model into the resampled training data, and lastly, I obtain my performance metrics by looking at my original, not resampled, test data. These steps should look familiar to you.8. Let's practice!
So, let's practice!Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.