Feature Selection and Classification of Spam on Social Networking Sites

Social networking sites (SNSs) see a variety of spam and scams targeted at their users. In contrast to the limited amounts of information available beyond message text and headers when analyzing email spam, spam on SNSs is often accompanied by a wealth of data on the sender, which can be used to build more accurate detection mechanisms. We analyze 4 million private messages as well as other public and private data from a popular social network in order to gain insight into the various features of spam messages and the accompanying user accounts data available to site operators. We use these insights to choose features that best differentiate spammers from legitimate, “ham,” users. Finally, we extract these features from the site’s data and use them to train and evaluate classifiers.