International Workshop on Affective Interaction between Humans and Machines in Multicultural Society (IWAM2022)

This topic aims to show a machine learning model that enables to predict signs of financial statement frauds by combining the domain knowledge of machine learning and accounting. Inputs of this model is a published dataset of financial statements, and outputs involve the conclusions whether the predicted financial statements indicate the signs of financial statement frauds or not. Currently, XGBoost is recognized as one of the most popular classification methods with fast performance, flexibility, and scalability. However, its default properties are not suitable for fraudulent detecting of imbalanced datasets. To overcome this drawback, this research introduces a new machine learning model based on XGBoost technique, called f(raud)-XGBoost. The proposed model not only inherits XGBoost advantages but also enables it to detect financial statement frauds. We apply the Area Under the Receiver Operating Characteristics Curve and NDCG@k to perform the evaluation process. The experimental results show that the new model performs slightly better than three existing models including logistic regression model that is based on financial ratios, Support-vector-machine model, and RUSBoost model.