Deep Feature Extraction Method for Automatic Classification and Processing of Accounting Information

In the context of advancing intelligent accounting systems within the broader field of computational methods and data processing, this study focuses on overcoming critical limitations inherent in conventional accounting information classification methodologies. These traditional systems often exhibit low adaptability to rapidly evolving data characteristics and a heavy dependence on manually crafted feature engineering pipelines. Existing approaches predominantly rely on predefined statistical heuristics or shallow machine learning models, which are insufficient for capturing the intricate, non-linear, and high-dimensional relationships embedded in modern financial datasets—especially those generated by large-scale enterprise resource planning systems. To address these constraints, we propose a novel deep feature extraction framework grounded in convolutional autoencoder architectures. This framework is designed to autonomously learn multi-level hierarchical feature representations directly from raw journal entries, eliminating the need for domain-specific feature selection. Our approach incorporates a dual-objective design that simultaneously performs input reconstruction and classification, allowing the network to preserve input fidelity while promoting discriminative learning. Furthermore, an adversarial training component is introduced to enhance generalization performance, especially under conditions of class imbalance and input noise—both common issues in financial transaction data.We evaluate our method on a comprehensive real-world accounting dataset composed of annotated journal logs across multiple business units. Experimental results demonstrate substantial improvements in both classification accuracy and anomaly detection F1-score when compared to traditional baseline models such as logistic regression and decision trees. Our findings illustrate the potential of deep learning-based architectures for intelligent accounting systems and contribute to the growing body of research on adaptive, scalable, and automated solutions for financial data analysis and auditing in computational finance.