add_new_oversampler.rst

Adding a new oversampler

Implementing a new oversampling logic which can be used in the model selection framework is easy:

• It should inherit from the smote_variants.OverSampling class
• Implement the class-level method parameter_combinations, which returns a list of reasonable parameter combinations compatible with the constructor. A parameter combination in the list needs to be a dictionary which can be passed to the constructor of the object using the asterisk-operator.
• It needs to implement the sample function, which takes a feature array and a target array.
• Finally, it needs to implement the get_params function to return the parameters of an oversampling instance as a dictionary.

Below can be found a template for adding new oversamplers:

class New_SMOTE_Variant(smote_variants.OverSampling):
    def __init__(self, param1, param2):
        super().__init__()

        self.param1= param1
        self.param2= param2

    @classmethod
    def parameter_combinations(cls):
        return [{'param1': 1, 'param2': 'a'},
                {'param1': 2, 'param2': 'b'},
                {'param1': 3, 'param2': 'c'}]

    def sample(self, X, y):
        # implement sampling logic here
        return X_samp, y_samp

    def get_params(self):
        return {'param1': self.param1, 'param2': self.param2}

An oversampler like this should work flawlessly with the model selection and evaluation scripts provided.