A Unified Hybrid Machine Learning Architecture for Robust Identity Anomaly Detection in Large-Scale Digital Ecosystems

Abstract

Identity anomaly detection in large-scale digital ecosystems is challenged by non-stationary behavioral dynamics, high-dimensional heterogeneous feature spaces, and limited availability of labeled anomaly instances. Existing approaches rely on isolated supervised or unsupervised models, restricting their ability to simultaneously detect known attack signatures and emerging behavioral deviations.This work introduces a unified hybrid learning architecture that addresses these limitations by integrating supervised classification, reconstruction-based unsupervised modeling, and temporal representation learning within a single optimization framework. By jointly optimizing discriminative and generative objectives through sliding-window aggregation and ensemble decision fusion, the architecture captures both short-term behavioral fluctuations and long-term identity evolution patterns. Empirical evaluation on large-scale identity interaction datasets demonstrates that the proposed framework achieves 97.3% accuracy and a 97.1% anomaly detection rate, outperforming strong supervised and unsupervised baselines by up to 7.7% in ADR. These results indicate that multi-paradigm temporal representation learning substantially enhances robustness to previously unseen anomalies under non-stationary conditions, providing a scalable foundation for identity-centric anomaly detection in complex digital environments.