AayushSameerShah / SMOTE

This small repository contains the SMOTE implementation from scratch.

SMOTE

This class is the implementation of SMOTE technique of oversampling the data points in the imbalanced dataset.

🗃️ Which files?

In this repository, 2 files are there.

1. .ipynb which is the step by step explanation of my approach
2. .py which is the standalone SMOTE class for direct use.

🤔 How to import?

This is not the library but instead a direct implementation. So where erver you want to use the SMOTE class, you can simply copy the file there and then follow the standard import syntax.

Here I have implemented in the simplest way possible and is dependent upon only numpy and sklearn's linear regression.

ℹ️ About

After you detect that you have got the problem of imbalanced instances per classes, you can simply use this class with ease.

After initializing, SMOTE class's object also has __repr__ which will display useful information.

🤓 Parameters

While initializing:

X: DF / ndarray / list
    It must be 2D and numerical in nature.

y: Series / array / list
    It must be 1D and should have categories in it.

oversampling_size: float (between 0 to 1)
    It ensures that how much percentage of highset frequencied class
    you want to generate new samples for. 1 means 100% which means
    total frequences of all class will be same.

While Resampling:

k: int
    This defines how many neighbours to choose while generating a new
    data point. Default is 5.

👨‍💻 How To

# Seperate the Features X and Labels Y
# X should be numerical and 2D.
X = df[["feature1", "feature2"]]
y = df["target"]

synthesizer = SMOTE(oversampling_size=0.5) # means 50% samples generated.
synthesizer.resample(k=3)

# ↓ this will return a dict with new data points
new_data = synthesizer.coords_by_class

Thanks!
Aayush ∞ Shah