STF: Steady and Transient Factorization for Sparse Time-Aware QoS Prediction

Abstract

Accurately predicting Quality of Service (QoS) under real-world sparsity is challenging because existing methods rarely distinguish between long-term steady trends and short-term transient perturbations, and often underutilize the collaborative filtering signals embedded in historical observations. To address these issues, we propose Steady and Transient Factorization (STF), a deep factorization framework that explicitly decomposes QoS into two interpretable components: a steady baseline capturing stable, long-term patterns, and a transient residual modeling short-term, interaction-specific variations. STF formulates time-aware QoS prediction as a multi-perspective learning problem, encompassing four complementary sub-tasks: user-wise, service-wise, user–service pair-wise, and user-service-time triplet-wise predictions, with dedicated steady and transient predictors applied to the first three. The outputs of these predictors are adaptively fused via a learnable weighting mechanism for end-to-end optimization, enabling the model to capture multi-level interaction granularity while enhancing robustness to sparsity. Extensive experiments on the large-scale dataset (30 million records) demonstrate that STF consistently outperforms state-of-the-art methods, reducing MAE and RMSE by up to 6.73% and 6.68% for response-time prediction, and reducing MAE and RMSE by up to 11.10% and 9.44% for throughput prediction, while maintaining competitive efficiency.