Semantic Graph Convolution Networks based Deep Learning Model for Spam Review Identification

Consumers heavily depend on users' evaluations of service-oriented websites. Numerous evaluations on review platforms, nonetheless, fail to accurately represent the genuine attributes of products due to business motivations. Consequently, monitoring website operations and maintaining a clean network environment requires the identification and management of fraudulent reviews. A semantic graph convolution network integrated with semantic similarity is introduced to improve the detection of erroneous research through text classification utilizing a graph neural network based on words and documents. Employ the TF-IDF feature value, the correlation between words and comments as determined by the PMI (Pointwise Mutual Information) index, and semantic similarity assessed by word embedding metrics. Upon the creation of the vocabulary-comment heterogeneous text network, the high-order feature information between word and comment nodes can be recorded and consolidated for classification purposes. The suggested strategy enhances accuracy by 7%, 4.8%, and 1.3% relative to CNN, LSTM, and Text-GCN on the public dataset.