RGF-Bot: A Novel Feature Selection Method to Identify Malicious Bot Accounts on Social Networking Sites Using Machine Learning

Abstract

A bot is an automated code used for malicious activities such as posting fake news, spreading malware, commenting on tweets, and liking the tweets, on Social Networking Sites (SNS) like Twitter. This paper proposes a novel feature selection method using machine learning to identify malicious bot accounts on social networking sites. This would help identify bot SNS accounts with minimal features yet maintain the same or higher accuracy. At the initial stage, the standard datasets from the Twitter platform were downloaded and pre-processed. Dataset 1, with 29 features and Dataset 2 with 30 features, were considered. The existing feature selection methods such as Variance Score (VS), Random Forest Importance (RFI), and Gradient Boost Importance (GBI) were applied to rank the features. Later, the proposed Recursive Grouping of Features (RGF) method is applied to VS, RFI, and GBI ranked feature sets to obtain the Minimal Features Sets (MFS)s in which the number of features is less than the total number of features. All classification algorithms were applied on VS, RFI, and GBI ranked MFSs to find the best-performing classifier and best feature ranking method. As a result, Decision trees were found to be the best classification algorithm on VS ranked MFSs. The proposed RGF method with the first MFS alone achieved the same accuracy on Dataset 1 and improved accuracy on Dataset 2 compared to all features.