Spotting Spammer using Group Spammer Behaviour Analysis

Abstract

Online shopping has become a metamorphic phenomenon in the IT world. Online reviews are considered the best source of customer’s opinion about a product, and an asset for customers, and organizations for making important buying decisions. Unfortunately, in order to increase the number of profits, promotion or even to demote a rival product, deceptive reviews (Spam) mislead prospective customers to buy the best product and organizations in decision making. Works have been proposed on detecting individual spammer reviews. But group review spamming, which includes a group of swindlers working together to post fake online reviews for promoting or demoting a product/s, has become more damaging. More the size of the group more difficult it is to differentiate them as fake reviews. At first, the work uses hints from behavioral data (timestamp, rating) and interpersonal data (network) to construct a suspicious reviewer graph. Then, it breaks the whole suspicious reviewer graph into k-clique clusters, and we consider such k-clique clusters as highly suspicious candidate group spammers. Finally, it ranks candidate groups by group based and individual-based spam indicators. Count Vectorizer and TF-IDF Vectorizer were used for results evaluation. TF-IDF Vectorizer gave the best results we have seen by providing us with the most reliable curve and greater AUC better than any other test cases. Random Forest gave us the best model with TF-IDF Vectorizer using k=7 and Unigram + Bigram + Trigram with Precision, Recall, and AUROC being 1, 1, and 0.74 respectively.