BTextCAN: Consumer fraud detection via group perception

Abstract

Traditional consumer fraud detection usually relies on the relevant regulatory authorities to conduct inspections through sampling. This would be labor-intensive and inefficient. To address this issue, we conducted a statistical analysis to explore the relationship between frauds and consumer perceptions. Based on the statistical results, we propose a novel deep mixture model-based consumer fraud detection method BTextCAN to detect consumer frauds via the perception of consumer group. By designing a text convolutional attention network (TextCAN) to extract local features with contextual semantic relations from consumer reviews, our approach can mine the opinions of consumers and use their group perception to detect consumer fraud behaviors. Experimental results show that our method outperforms the baseline models. In particular, BTextCAN achieves an accuracy of 79.8% in the binary detection task and 76.5% in the multiclassification detection task. This work is the first research effort to detect fraudulent merchant behavior from consumer reviews. In addition, we have collated and made publicly available the first dataset in this area.