Hybrid awred: synergy of adaptive reconstruction and topological clustering for anomaly detection in multimodal data

DOI: 10.31673/2412-9070.2026.017405

Abstract

The rapid digitalization of the financial sector and the growth of transaction volumes intensify the challenge of automated fraud detection. Anomaly Detection in modern data streams is characterized by two fundamental problems that complicate the application of classical algorithms: extreme class imbalance (the proportion of anomalies is often less than 0.1%) and the complex multimodal structure of clients' normal behavior. Traditional deep learning methods demonstrate limited effectiveness under such conditions. In particular, autoencoders (AE) and their variations are prone to overfitting on the majority class, minimizing the average error at the expense of ignoring rare events. At the same time, one-class classification methods, such as Deep SVDD, are effective for unimodal data; however, they destroy the local topology of multimodal distributions by attempting to collapse hetero-geneous clusters of normal data to a single hypersphere center, leading to the masking of anomalies.
This paper presents a novel method, Hybrid AWRED (Adaptive Weighted Reconstruction with Regularized Energy and Dynamics), developed for robust anomaly detection in complex environments. The proposed approach implements the synergy of three dynamic mechanisms for the first time. First, a Self-Weighted Error Feedback mechanism is introduced, which automatically focuses the model's attention on difficult examples without the need for synthetic data generation. Second, a hybrid loss function has been developed, combining a modified “Center Loss” for cluster compactification and topological variance stabilization to prevent latent space collapse. Third, a key innovation is the use of an oscillating regularization coefficient that dynamically shifts the priority between pre-serving data structure (Manifold Learning) and compressing it, allowing the model to iteratively escape local minima.
Experimental evaluation conducted on the synthetic “Hard Mode Credit Card Fraud” dataset (60,000 records, 41 features) confirmed the superiority of the proposed architecture. Hybrid AWRED achieved an AUC-ROC of 0.9873 and a Recall of 0.7043. Comparative analysis demonstrated that the method outperforms the SOTA algorithm Deep SVDD by 35% in the critical metric of detecting hidden attacks, ensuring a better balance between sensitivity and specificity. The obtained results open new perspectives for building reliable unsupervised financial monitoring systems.

Keywords: deep learning; anomaly detection; Hybrid AWRED; Deep SVDD; Center Loss; adaptive regularization; imbalanced data; multimodal distributions.