STAR: A Sequential Transformer-Based Adaptive Reinforcement Learning Framework for Service Recommendation

Service requirements in service computing often exhibit semantic decomposability, where complex demands can be fulfilled progressively through multiple functional aspects. Motivated by this observation, We propose STAR, a Sequential Transformer-based Adaptive Reinforcement Learning framework. STAR adopts a fulfillment-oriented modeling paradigm that reformulates service recommendation as a progressive demand-satisfaction process rather than a one-shot matching task. It integrates graph-based service representation learning with a residual-requirement updating mechanism to capture both general requirement-related semantics and the dynamically preserved unmet semantics in complex service demands. To further support stable sequential recommendation, STAR employs a two-stage optimization strategy with warm-up, feedback-driven refinement, and retained ground-truth supervision to mitigate requirement drift caused by progressive demand updating. Experiments on the ProgrammableWeb dataset show that STAR outperforms competitive baselines across multiple evaluation metrics, demonstrating its effectiveness.