An Research On Handling User Cold-Start Problem In Review Spam Detection

In recent years,E-commerce has come to pervade numerous aspects of our lives with great convenience.Online reviews record the sincere experiences of real reviewer,and have become an important reference for consumers to make consumption decisions.The impelling of enormous commercial benefits results in the emergence of spam reviews.Obviously,spam reviews will seriously mislead consumers' choices and cause negative consumption experience,subsequently affect the development and reputation of business and review website.Hence,the detection of fake reviews has become an urgent research problem.Previous studies are mainly focused on extracting effective linguistic or behavioral features to distinguish the spam and legitimate reviews.Such features are either ineffective or take long time to collect and thus are hard to be applied to user cold-start spam review detection tasks.There is few research on user cold-start spam review detection.Recent advance leveraged the neural network to encode the textual and behavioral features to alleviate the data sparcity caused by user cold-start problem.However,the abundant attribute information are largely neglected by the existing framework and the performance are not promising enough.In this paper,we first propose a BFGN model to generate the behavioral features for cold-start review.With in-depth analysis of the weakness of existing methods and our BFGN model,we propose AE model to incorporate entities and their inherent attributes into a unified framework.Specifically,our AE model not only encodes the entities of reviewer,item,and review,but also their attributes such as location,date,price ranges.Furthermore,we propose a domain classifier to adapt the knowledge from one domain to the other,and improve AE to AEDA model.With the abundant attributes in existing entities and knowledge in other domains,we can successfully cope with the problem of data scarcity in the user cold-start settings.Experimental results on two Yelp datasets prove that our proposed framework significantly outperforms the state-of-the-art methods.