Impact of Behavioral and Textual Features on Opinion Spam Detection

Abstract

Opinion spam or fake review detection is extensively a widespread and established problem. Due to this, it has received a substantial amount of research attention in the past decade. The prime actors responsible to orchestrate a spamming situation are commonly known as opinion spammers. Such individuals aim at manipulating online review platforms by either promoting or demoting the reputation of a target product or business, thereby misleading the consumers. By considering opinion spam as a classification problem, a variety of spamming features viz., behavioral, textual, network-based etc., have been engineered and utilized in the literature. However, examining the effectiveness of these features has not been studied so far. This paper aims at investigating the impact of behavioral and textual features in connection to both review-centric (spam detection) and reviewer-centric (spammer detection) settings. For the same, support vector machine (SVM) models have been trained on YelpZip dataset. Our results show that behavioral features comprehensively outperform textual features in detecting both spam and spammers. Additionally, the posterior analysis performed on both feature sets also confirms the same.