A Region-Aware Dual Latent State Mining Framework for Service Recommendation in Large-Scale Service Networks

In large-scale service networks, Quality of Service (QoS) data is vital for tasks such as resource provisioning, real-time service recommendation, and user experience optimization; yet effectively predicting QoS often faces two key obstacles: data sparsity and highly imbalanced distributions (where most response times cluster near small values while a minority grow disproportionately large). Existing approaches typically rely on individual object (user/service) features, overlooking the phenomenon that users or services within the same region (physical or virtual) exhibit similar network states. This paper presents a Region-Aware Dual-Latent State Learning (R2SL) framework that tackles these challenges by explicitly modeling regional network latent states. Specifically, we propose to learn physical-region (city-level) and virtual-region (AS-level) latent states from historical QoS records through a joint EM–gradient descent strategy, thereby alleviating data sparsity. Furthermore, to mitigate label imbalance in QoS data, we introduce a Smooth Huber (S-Huber) loss function that appropriately reweights extreme errors, preventing the training process from being dominated by outliers. We also develop a sparsely activated mixture-of-experts module, dynamically routing regional latent features based on each prediction task’s context. Experiments on real-world QoS datasets show that R2SL substantially outperforms state-of-the-art baselines, including the newly introduced FRLN. On throughput tasks, R2SL reduces MAE by an average of 18.8% and RMSE by 12.9%, while on response time tasks, it achieves 26.4% lower MAE and 24.2% lower RMSE. These findings indicate that dual-latent state modeling, combined with a distribution-aware loss, effectively captures complex regional patterns and mitigates long-tail label effects, making R2SL a powerful and scalable framework for large-scale QoS data mining in service networks.