A New Ensemble Model for Phishing Detection Based on Hybrid Cumulative Feature Selection

A New Ensemble Model for Phishing Detection Based on Hybrid Cumulative Feature Selection(PDCFS) proposes a model that partitions the main dataset into
$n$
partitions based on the features available in the dataset. This process is done by feeding the dataset into multiple feature selection methods: Chi-Square, Gain Ratio, Information Gain, Pearson Correlation Coefficient, and Principal Components Analysis, and arranging the dataset into
$n$
datasets taking top-n features including the class label, given by the filter method, and discarding the remaining ones each time. Then each partitions of the dataset are used for training and testing using 5-fold cross validation, applying a number of classifiers: Support Vector Machine, Naïve Bayes, C4.5, Random Forest, JRip, PART, and k-Nearest Neighbors. In the next, the results are voted to get the best possible result. Majority voting is applied on both reduced feature subsets gained through feature selection steps, and full feature set to draw a comparison between the feature sets followed. The overall speculation suggests that Random Forest with reduced feature set of 32 tops the result chart with and accuracy of 98.36 %, while proposed PDCFS scores 98.24 % of accuracy. PDCFS also demonstrates a comparative performance when compared with other hybrid models.