Abstract
A forecasting architecture is described herein that uses a global Light Gradient Boosting Machine (LightGBM) trained exclusively on synthetic data parameterized by contact-center queue-behavior archetypes and operating alongside a lazy k-nearest neighbor (KNN) baseline. Both models independently predict each interval, with the lazy learner serving as a calibration anchor. A feedback loop compares the eager model output with the lazy baseline, extracts pattern characteristics of missed intervals and calibration gaps, and uses those characteristics to refine synthetic data generation and feature engineering for a later retraining cycle. Independent quality gates for call volume and average handle time (AHT) determine whether a queue uses the eager output, a hybrid forecast, or the lazy baseline. The eager model also consumes a Rolling Adaptive Quantile Score Selection (RAQS) mechanism for contextually optimal quantile selection. The result is a progressively improving eager learner that captures cross-queue structural patterns without exposing raw customer data.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Vaidyanathan, Iyer Venkataraman; Kumar, Dinesh S.; Mohamed, Dhaanish; B A, Sudheendra; Daulath, Rajiv Shankar; Biswas, Supratik; B M, Hemanth Kumar; and Banerji, Nikita, "SYNTHETIC-ONLY GLOBAL GRADIENT BOOSTING WITH EAGER-LAZY COMPLEMENTARY FORECASTING AND FEEDBACK-INFORMED RETRAINING FOR CONTACT-CENTER WORKLOAD", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12217