Opinion spam detection by incorporating multimodal embedded representation into a probabilistic review graph.

Spam reviews typically appear perfectly normal until examined in a large context. The standard approach to classifying reviews independently ignores these relations. In this study, we propose a complex probabilistic graph classification approach to address the problem of opinion spam detection. To obtain an initial effective spamicity estimation for the nodes (reviews, authors, and products) in the graph, we first train a neural network with attention mechanism to learn the multimodal embedded representation of nodes by leveraging both textual and rich features. Then based on the node prior computation, a heterogeneous graph is constructed to capture the relationships among different kinds of nodes, and the beliefs are further updated through iterative message propagation. To support this work, we collect two kinds of real-life datasets, which are separately composed of 97,839 restaurant reviews and 31,317 hotel reviews. The evaluation of the two datasets demonstrates the effectiveness of the proposed approach. We further analyze several salient rich features and the intermediate component of our model, thereby revealing that their states capture certain statistical characteristics of the datasets.