Analysis of Dataset balancing Techniques for Phishing Dataset

One of the major aspects that affects the performance of a machine learning based algorithm is the dataset. A huge dataset is not the only parameter that will improve the performance of the classifier but the balance between the classes for classification is also a crucial factor. An unbalanced dataset will be biased towards the prediction for majority class samples. The classifier trained on unbalanced dataset is more likely to predict false positive or false negative results for minority class samples. The problem of unbalanced dataset is observed in phishing dataset since the number of genuine web URLs is more as compared to the fake URLs. This paper discusses 8 different techniques like SMOTE, SMOTE+Tomek, SMOTE+OSS, SMOTE+ENN, ADASYN, ADSYN+Tomek, ADASYN+OSS, ADASYN+ENN that can be used to generate a balanced dataset and evaluates the them based on the performance of KNN and Random forest classification models that are trained using the original and balanced dataset.