AIS-Oversampler

This is an Oversampler for handling Class Imbalance in Binary Classification tasks. It uses the AISOv algorithm that is based on Artificial Immune Systems.

The oversampler works for binary classification and can be set to use different mathematical modes of operation that alter how the final resampling data is generated. The AISOv algorithm has been tested on a variety of datasets and performs comparably to other common and proven oversampling techniques.

Installation

The AIS Oversampler requires:

• Python (>= 3)
• Scikit Learn (>= 1.2)
• Pandas (>= 1.4)

How to use it

This is a simple code example:

from AIS_Oversampler import ArtificialImmuneSystem 

#create an instance of the oversampler
oversample_AIS = ArtificialImmuneSystem.ArtificialImmuneSystem()

#import a dataset
df = pd.read_csv("../datasets/dataset2.csv",index_col=0)

#separate the features and labels
features = df.drop(["5"], axis=1)
label = df.drop(df.columns[0:-1],axis=1)

#Initialize a classfier
randomForest = RandomForestClassifier()

#call the main function AIS_Resample with the required parameters
df_after_oversampling = oversample_AIS.AIS_Resample(features, label, model = randomForest)

#call the main function AIS_Resample with the required parameters and optional parameters
df_after_oversampling = oversample_AIS.AIS_Resample(preparedDF, labels, max_rounds = 50, stopping_cond = 20, model = randomForest ,K_folds = 5,scorer = 'f1',min_change = 0.005, use_lof = False, mutation_rate = 1.0)

Main Function Parameters

Note: Optional Parameters have a default value but should be given a value and fine tuned for optimal results

How it works

The AISOv algorithm takes in an Imbalanced dataset and preprocesses it to determine how many antibodies to create and the necessary parameters for the other parts of the function. It then creates an Initial antibody population before entering the Main Program Loop. In the loop, the antibody population is mutated, a fitness function and an evaluation function are applied to find the current best antibody population. Once the termination condition for the loop is met, the current best antibody population is returned.

You can read more about it here: https://docs.google.com/document/d/1JGAjYWz2Wp95ArWZQnXeLTWyc3ArYzahvMBh8H3XbPA/edit?usp=sharing

About

This algorithm was originally developped by myself, Nikhil Pyndiah alongside my teamates Adam Jansen and Jacob King for our Honours project. Our aim with the project was to create an easy to use oversampling algorithm which could be used as a drop in replacement for other similar oversamplers. The original code alongside extensive testing we did can be found in the report I linked to above and the original code can be found here: https://github.com/nikhil815/Artificial-Immune-System-For-Class-Imbalance