A Novel Cost-sensitive Capsule Network for Audit Fraud Detection

In the face of increasing audit tasks, it is imperative to realize the transformation from manual audit based on domain expertise to intelligent audit based on algorithm, in order to improve the efficiency and quality of audit. However, the serious imbalance of data distribution and strong dependence on domain experts pose a huge challenge to the application of the algorithm in the field of audit. In response to the above challenges, this paper proposes a cost-sensitive capsule network (CSCN) to identify abnormal accounting vouchers. It can not only successfully extract features from multiple types of input data and improve the accuracy through capsule network, but also deal with extremely unbalanced data by introducing cost-sensitive loss function. The comparative experiment results show that the proposed CSCN algorithm accurately identifies all abnormal accounting vouchers, the G-mean index is increased by 5.3% and the average cost index is reduced by 64.6%, which fully verifies its effectiveness.