Design of multi-task learning model for anomaly detection and risk early warning of transactional accounting data
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The characteristics of transactional accounting data, such as high dimension, strong time series and high correlation, lead to the problems of task separation, rule solidification and poor adaptability in the traditional anomaly detection and risk early warning model. This article proposes a multi task learning (MTL) framework that integrates business rules and deep learning. For the first time, anomaly detection and risk warning are modeled together. The underlying bidirectional LSTM network (Bi-LSTM) is shared to extract temporal and business features, and anomaly recognition and future risk prediction are achieved through task specific headers. The weighted loss function and rule regularization term are designed, taking into account sample imbalance and criterion compliance, and supporting online incremental learning to cope with data distribution drift. Experiments based on 5 million pieces of real enterprise desensitization accounting data show that the model is significantly superior to single task and traditional model in anomaly detection F1-score of 0.681, recall rate of 0.835, and risk early warning F1-score of 0.723, with an average of 7.5 days early warning. Ablation experiment and visual analysis verify its synergistic effect, interpretability and robustness, and provide credible and traceable intelligent support for enterprise internal control and audit.
Fields of Study & Subjects
Anomaly Detection Techniques and Applications
Financial Distress and Bankruptcy Prediction
Imbalanced Data Classification Techniques
Publication details
- Journal / source
- IET conference proceedings.
- Volume
- 2025
- Issue
- 47
- Pages
- 786–790
- Publication date
- 1 March 2026
- Language
- EN