Anomaly detection in the edges of dynamic graphs has become one of the central challenges in data science an‎d machine learning due to the unprecedented growth in data volume an‎d complexity in applications such as fraud detection, social network monitoring, an‎d cybersecurity. The need for real-time processing an‎d memory constraints, combined with the lack of precise labels in many real-world datasets, makes the direct use of heavyweight, supervised algorithms problematic. In this study, we propose a lightweight, distributed, an‎d modular framework to address this issue, consisting of three main components: (1) data distribution an‎d windowing to reduce memory footprint an‎d facilitate parallel processing; (2) anomaly scoring using the AnoEdge-G algorithm; an‎d (3) an automated thresholding module incorporating a set of statistical methods alongside an innovative approach based on a multiple of the maximum scores within each time window. Employing static windowing not only significantly reduces the processing load but also enables this scheme to generalize to future data streams while maintaining efficiency in real-time scenarios. All stages of the framework are designed to be deterministic, reproducible, an‎d fully parallelizable Initial experiments assessing the impact of parallelization on execution time an‎d accuracy degradation demonstrated that, with appropriate data partitioning, processing speed can be dramatically increased with only a minor dro‎p in accuracy—making the framework highly suitable for scalable, real-time applications. During threshold determination, our novel method performed at least as well as established techniques an‎d, despite its simplicity, achieved comparable accuracy in anomaly identification. In one of the benchmark datasets, anomaly detection accuracy reached approximately 99%, indicating a significant improvement offered by the proposed framework. These results demonstrate that our approach represents an effective step toward operationalizing anomaly detection in dynamic graphs under resource constraints an‎d real-time processing requirements.

عليرضا بصيري

زينب مالكي , زينب زالي