A Predict-then-Optimize Job Assignment Framework for Efficient Trademark Application Evaluation

Timely evaluation of trademark applications is critical for effective intellectual property protection, yet many national offices face rising workloads and limited staffing flexibility. At TURKPATENT, these pressures have contributed to growing backlogs and high variability in evaluation times. This study proposes a data-driven job assignment framework that integrates machine learning based prediction of examiner-specific completion times with a rolling-horizon mixed-integer programming model. A comprehensive feature set incorporating text indicators, classification codes, examiner effects, and workload characteristics is used to train predictive models, and predicted durations are subsequently embedded into an assignment model designed to minimize tardiness while maintaining workload balance. Multiple assignment policies are evaluated through extensive simulation under varying workload and prediction-quality scenarios. The results show that the predictive rolling-horizon model reduces tardy jobs by roughly 49.5% relative to the current system. A fully automated operational pipeline was implemented to enable daily deployment, and a real-world pilot with eight examiners achieved a 35.3% reduction in tardy jobs. The study demonstrates that integrating predictive analytics with optimization can substantially improve performance in administrative workflows and offers a scalable approach for modernizing trademark examination processes.