Abstract
With the advent of social networking sites, opinion-mining applications have attracted the interest of the online community on review sites to know about products for their purchase decisions. However, due to increasing trend of posting spam (fake) reviews to promote the target products or defame the specific brands of competitors, Opinion Spam detection and classification has emerged as a hot issue in the community of opinion mining and sentiment analysis. We investigate the issue of Opinion Spam detection by using different combinations of entities, features, and their sentiment scores. We enrich the feature set of a baseline Spam detection method with Spam detection features (Opinion Spam, Opinion Spammer, Item Spam). Using a dataset of reviews from the Amazon site and sentences labeled for Spam detection, we evaluate the role of spamicity-related features in detecting and classifying spam (fake) clues and distinguishing them from genuine reviews. For this purpose, we introduce a rule-based feature weighting scheme and propose a method for tagging the review sentence as spam and non-spam. Experiments results depict that spam-related features improve Spam detection in review sentences posted on product review sites. Adding a revised feature weighting scheme achieved an accuracy increase from 93 to 96%. Furthermore, a hybrid set of features are shown to improve the performance of Opinion Spam detection in terms of better precision, recall, and F-measure values. This work shows that combining spam-related features with rule-based weighting scheme can improve the performance of even baseline Spam detection method. This improvement can be of use to Opinion Spam detection systems, due to the growing interest of individuals and companies in isolating fake (spam) and genuine (non-spam) reviews about products. The outcome of this work will provide an insight into spam-related features and feature weighting and will assist in developing more advanced applications for Opinion Spam detection. In the field of Opinion Spam detection, previous state-of-the-art studies used less number of spamicity-related features and less efficient feature weighting scheme. However, we provided a revised feature selection and a revised feature weighting scheme with normalized spamicity score computation technique. Therefore, our contribution is novel to the field because it provides a significant improvement over the comparing methods.