On Feature Selection for the Prediction of Phishing Websites

with the rise of the big data paradigm, large data sets are being made available for knowledge mining. While this open up possibilities for new insights being gained every day, it also exposes data consumers to an increase in low quality, unreliable, redundant or noisy portions of the data. This would negatively affect the process of harvesting knowledge and recognizing patterns. Therefore, efficient feature selection methods to empower for real-time prediction or classification systems. Feature selection is the process of identifying the most relevant attributes and removing the redundant and irrelevant attributes. In this study, we implemented Kaiser-Meyer-Olkin (KMO) Test as a feature selection method and applied that to a publicly available phishing dataset, namely, the UCI of phishing website. furthermore, we used Logistic Regression and Support Vector Machine as classification methods to validate the feature selection method. Our results show just a slight difference in accuracy between implementation using full dataset features and the proposed much smaller dataset (almost 63% of original features set). This reduction in dimensionality is significant for the realtime systems especially when the accuracy reduction is slight. From there, we present a framework enabling a significant reduction in features. This opens the door for future work under which a wider set of classification algorithms will be tested in order to achieve the dimensionality reduction and an increase in performance accuracy.