Deep feature fusion for cold-start spam review detection

The cold-start problem in spam review detection is a significant challenge referring to identifying the authenticity of the first review posted by new users. For generating more sensitive features to identify new reviews, existing methods mainly leverage text-similarity of review to find relevant features to approximate the incomplete behavior features of new reviews. However, they over-rely on the text information of new reviews while ignoring the mutual behavioral information in the review system, leading to a decrease in the sensitivity of features. To address the issue, we propose a deep feature fusion method, which balances the importance of text information and behavior information to enhance features’ sensitivity. Specifically, we construct a heterogeneous graph, where products and users serve as vertices connected by edges representing reviews. Then, we perform graph convolution calculation on this graph in the first feature fusion stage. We utilize the mutual behavioral information in the review system to compensate for the incomplete behavior feature of new reviews. Furthermore, we design a co-attention network, which can give features different weights in the global feature fusion stage, to gain features with high sensitivity of identifying new reviews. Extensive experiments on Yelp-hotel and Yelp-restaurant datasets demonstrate that our proposed approach yields better classification performance over existing methods.