Anomaly-resilient and robust learning frameworks for real-time QoS prediction

Abstract

Quality of Service (QoS) prediction plays a crucial role in ensuring reliable and efficient operation of modern distributed digital services. Accurate estimation of performance attributes, such as response time, throughput, reliability, and availability, is essential for maintaining service-level guarantees, optimizing resource allocation, and enhancing user experience. These attributes directly influence system performance and user satisfaction. In many applications, including cloud platforms, edge computing, and intelligent transportation systems, accurate QoS estimation is critical for ensuring stable, efficient, and safe system operations. Despite significant research progress, accurate QoS prediction remains challenging due to the inherent characteristics of real-world service interaction data, particularly extreme sparsity. In practice, the user–service interaction matrix is highly incomplete, as users typically invoke only a small subset of available services. As a result, observed QoS entries represent only a limited portion of all possible interactions, restricting reliable supervision and weakening correlation learning. This sparsity significantly limits model generalization to unseen or infrequently invoked services, complicating the development of robust and reliable QoS prediction frameworks.