Local interpretable spammer detection model with multi-head graph channel attention network

Fraudulent reviews posted by spammers on the online shopping websites mislead consumers' purchasing decisions. To curb fraudulent reviews, many methods have been proposed for detecting spammers. However, the existing spammer detection methods operate in a "black box" mode and lack a reasonable interpretation for their detection results. Aiming at this concern, we propose a local interpretable spammer detection model with multi-head graph channel attention network. First, we design a multi-head graph channel attention network to aggregate neighbor node features from different angles and employ the captured high-order features between users to detect spammers. Then, we collect interpretable evidences from the extracted features to interpret the detection results by combining the HSIC Lasso algorithm and random walk with restart strategy. Based on which, we select the important features that affect the detection results as the final explanations. Experiments indicate that improvements of accuracy, precision, recall and F1-measure of our model over the state-of-the-art solutions on the Amazon, YelpChi, YelpNYC, YelpZip, and Yelp_four datasets are [2.9 %, 1.98 %, 2.53 %, 4.37 %], [13.81 %, 8.91 %, 1.68 %, 9.55 %], [4.25 %, 16.13 %, 1.56 %, 15.13 %], [11.11 %, 8.55 %, 8.22 %, 11.75 %], and [4.46 %, 4.58 %, 7.9 %, 7.01 %], respectively. The frequency distributions of our interpretation results without noise features and the Fidelities of our interpretation method under different Sparsity levels are the best on these datasets.