Financial Feature Embedding with Knowledge Representation Learning for Financial Statement Fraud Detection

Abstract

Identifying fraudulent financial statement is critical for capital market regulation and is generally formulated as a classification problem. Feature selection in traditional machine learning methods does not consider correlation information among financial features which may influence performance of classifiers. To explore correlation information on conducting financial statement fraud detection (FSFD), we combine traditional features with knowledge graph models, and learn new representations enriched with feature embedding of various financial categories. These feature relations defined by correlation types may form knowledge graphs with features as nodes and correlation relations as edges. Experimentations demonstrate that financial feature representations with correlation information significantly improve classification performances for SVM and K-NN, marginally better than decision trees and logistic regression, but not outperforming naive Bayes (Kernel).