Aspect-based classification method for review spam detection

Online reviews have become available for consumers’ reference to make purchase decisions, but a large number of spam reviews have damaged e-commerce reputations. Previous research has addressed review spam detection with classification models using textual features, behavior features, and relational features. However, the fine-grained aspect features related to the product attributes in online reviews have been overlooked and have not yet been thoroughly studied. Therefore, this study proposes a review spam detection model based on a list of novel aspect features. The basic idea is that since spam reviews are usually written by users without real experience, the product aspects depicted in spam reviews will be different from those in genuine reviews. First, we use the Bi-LSTM model to automatically extract massive aspect words, which are then clustered into different aspect categories by the K-means algorithm. Further, we propose nine novel aspect features to train a machine learning model for review spam detection. Experimental results on two labeled Yelp datasets show that the proposed aspect features can significantly improve the accuracy of review spam detection by about 16.11% to 38.86% compared with textual and behavior features.