Cost-Sensitive Classifier for Spam Detection on News Media Twitter Accounts (revised April 2017)

Abstract

Abstract—Social media are increasingly being used as sources in mainstream news coverage. However, since news is so rapidly updating it is very easy to fall into the trap of believing everything as truth. Spam content usually refers to the information that goes viral and skews users views on subjects. Despite recent advances in spam analysis methods, it is still a challenging task to extract accurate and useful information from tweets. This paper aims at introducing a new approach for classification of spam and non-spam tweets using Cost-Sensitive Classifier that includes Random Forest. The approach consisted of three phases: preprocessing, classification and evaluation. In the preprocessing phase, tweets were first annotated manually and then four different sets of features were extracted from them. In the classification phase, four machine learning algorithms were first cross-validated aiming at determining the best base classifier for spam detection. Then, class imbalanced problem was dealt by resampling and incorporating arbitrary misclassification costs into the learning process. In the evaluation phase, the trained algorithm was tested with unseen tweets. Experimental results showed that the proposed approach helped mitigate overfitting and reduced classification error by achieving an overall accuracy of 89.14% in training and 76.82% in testing.