Generating Behavior Features for Cold-Start Spam Review Detection with Adversarial Learning

Abstract Due to the wide applications, spam detection has long been a hot research topic in both academia and industry. Existing studies show that behavior features are effective in distinguishing the spam and legitimate reviews. However, it usually takes a long time to collect such features and thus is hard to apply them to cold-start spam review detection tasks. Recent advances leveraged the neural network to encode the various types of textual, behavior, and attribute information for this task. However, the inherent problem, i.e., lack of effective behavior features for new users who post just one review, is still unsolved. In this paper, we exploit the generative adversarial network (GAN) for addressing this problem. The key idea is to generate synthetic behavior features (SBFs) for new users from their easily accessible features (EAFs). Specifically, we first select six well recognized real behavior features (RBFs) existing for regular users. We then train a GAN framework including a generator to generate SBFs from their EAFs including text, rating, and attribute features, and a discriminator to discriminate RBFs and SBFs. We design a new implementation of generator and discriminator for effective training. The trained GAN is finally applied to new users for generating synthetic behavior features. We conduct extensive experiments on two Yelp datasets. Experimental results demonstrate that our proposed framework significantly outperforms the state-of-the-art methods.