Sparsity-resilient QoS prediction via ε-DP enhanced subgraph-inductive GNNs in internet of services

Abstract

The rapid evolution of the Internet of Services (IoS) in the Web 3.0 era empowers users with access to an abundance of functionally similar services. This makes reliable Quality of Service (QoS) prediction essential for selecting optimal services. However, real-world QoS data often suffer from high sparsity, and privacy concerns frequently prevent users from sharing their raw QoS records. This creates a dual challenge: achieving high prediction accuracy while preserving data privacy in sparse environments. To address these challenges, we propose DPIS-GNN, a novel framework that integrates ε-Differential Privacy (ε-DP) with subgraph-inductive Graph Neural Networks (GNNs). Our approach first applies ε-DP mechanisms to obfuscate local datasets, protecting sensitive user information. These perturbed datasets are then aggregated into a unified interaction graph, from which relevant patterns are inferred. Next, we introduce a subgraph-based GNN that inductively learns from the noisy and sparse data to produce accurate QoS predictions. Extensive experiments on real-world datasets demonstrate the effectiveness of our approach. DPIS-GNN achieves up to a 17.55% reduction in Mean Absolute Error (MAE) and a 9.77% decrease in Root Mean Square Error (RMSE), outperforming state-of-the-art baselines. Our model exhibits superior robustness in sparse and cold-start scenarios, offering both strong privacy protection and high predictive performance.