FIGDA: Feature Importance Guided Learning Model for Domain-Adaptive Anomaly Detection

Recent advancements in machine learning, particularly deep learning, have substantially improved the analysis of high-dimensional data. However, inter-domain heterogeneity remains a key challenge that undermines the reliability of anomaly detection. As a result, models that perform well on one dataset often underperform on others even after retraining with domain-specific data, owing to inter-domain discrepancies in feature characteristics and anomaly patterns. To address this issue, this paper proposes Feature Importance Guided Learning Model for Domain-Adaptive Anomaly Detection (FIGDA), an extension of Deviation Networks (DevNet) that integrates feature importance (FI) extracted from Extreme Gradient Boosting (XGBoost) to guide both weight initialization and optimization in the neural weight space. The proposed model initializes the first layer of the network based on FI and introduces a regularization term in the loss function to reflect FI-guided constraints. We conducted experiments on four public benchmark datasets spanning demographic, financial, and visual domains. In the Census dataset, FIGDA achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.919, improving by 0.188 and 0.050 points compared to DevNet and the Transformer–CNN–based Anomaly Detection Network (TCAD-Net), respectively. In the Bank dataset, FIGDA reached an AUC-ROC of 0.931, exceeding DevNet’s 0.741 by 0.190 points. Moreover, our proposed FIGDA demonstrated consistent performance across ten runs on highly imbalanced datasets, confirming robustness in challenging real-world conditions. These results highlight that the integration of deviation-based modeling with FI helps the model adapt to domain-specific characteristics and yields stable performance.