Decoupling Location and Preference: A Dual-Branch Architecture for Robust QoS Prediction Under Extreme Sparsity

Quality of Service (QoS) prediction faces challenges from location-dependent variability and sparse user-service interactions. Existing methods often struggle to integrate location information (e.g., using fixed weights for spatial attributes) or learn representative features from sparse matrices. This paper proposes a method to Decoupling Location and Preference via a dual-branch architecture for robust QoS prediction under extreme sparsity, called DLP. It integrates location and preference features to address the challenges of sparsity and contextual variability. Unlike conventional single-stream or simple concatenation methods, DLP features a novel dual-branch architecture that decouples heterogeneous features and specializes in processing them: Location context and user-service preferences. The first branch, a location feature extraction network, processes user and service geographical and network information. It utilizes an attention mechanism to dynamically weight spatial attributes (instead of fixed weights) based on their actual impact on QoS and selects the most salient co-location features to model spatial interactions. The second branch, a preference feature extraction network, constructs high-dimensional feature representations from similarity-based user-service vectors derived from the sparse QoS matrix. It employs a multi-layer feature extraction block that hierarchically aggregates intermediate features to compensate for information loss during transformation, thereby capturing richer user/service preferences. Finally, a feature fusion prediction network integrates the learned location and preference features to generate accurate QoS predictions. Ablation studies and analysis validate that each component contributes significantly to performance gains. Extensive experiments on the WS-DREAM dataset show that DLP outperforms 22 baselines across 2.5%–20% sparsity, excelling in throughput prediction (achieving reductions up to 9.07% in MAE and 28.86% in RMSE at 2.5% sparsity) and validating its superior QoS prediction accuracy.