Streamlining the Detection of Accounting Fraud through Web Mining and Interpretable Internal Representations

摘要

Considerable effort has been devoted to the development of Artificial Intelligence tools able to support the detection of fraudulent accounting reports. Published results are promising but, till the present date, the use of such research has been limited, due to the "black box" character of the developed tools and the cumbersome input task they require. The tool described in this paper solves both problems while improving specificity of diagnostics. It is based on Web Mining and on Multilayer Perceptron classifiers where a modified learning method leads to meaningful representations. Such representations are then input to a features' map where trajectories towards or away from fraud and other features are identified. The final result is a robust Web Mining-based, self-explanatory fraud detection tool.