Multimodal anomaly detection framework in intelligent financial audit: based on LightGBM-SVM hybrid model and dynamic feature enhancement

In response to the complexity and industry-specific differences in detecting anomalies in financial data of listed companies, this study proposes a multimodal detection framework that integrates dynamic feature enhancement with hierarchical optimization. Based on a hybrid model combining Support Vector Machines (SVMs)and Lightweight Gradient Boosting Machines (LightGBM), the framework significantly improves the accuracy and efficiency of anomaly detection through industry-specific feature screening, collaborative training of a two-layer model, and dynamic parameter optimization. The experiment uses financial data from 2,667 listed companies over nearly six years, covering high-frequency industries such as manufacturing and information technology services. It extracts industry-specific indicators (such as "construction in progress volatility" and "R&D expense ratio") and global common features (such as "abnormal growth rate of cash and cash equivalents"), constructing a multimodal feature set. The first layer of the SVM model screens for high-confidence anomaly samples, while the second layer of the LightGBM achieves refined classification through histogram algorithms and feature fusion. Results show that the hybrid model outperforms traditional algorithms (such as random forests and logistic regression) in terms of accuracy (0.912), recall (0.954), and AUC value (0.980), particularly reducing the false positive rate in manufacturing to 4.2%. This framework provides an efficient and scalable solution for intelligent financial auditing, supporting real-time monitoring needs for high-dimensional data across multiple industries.