A Conceptual Model for Labeling in Reinforcement Learning Systems: A Value Co-creation Perspective

Artificial intelligence (AI) possesses the potential to augment customer service employees e.g. via decision support or solution recommendations. Still, its underlying data for training and testing the AI systems is provided by human annotators through human-in-the-loop configurations. However, due to the high effort for annotators and lack of incentives, AI systems face low underlying data quality. That in turn results in low prediction performance and limited acceptance by the targeted user group. Faced with the enormous volume and increasing complexity of service requests, IT service management (ITSM) especially, relies on high data quality for AI systems and incorporating domain-specific knowledge. By analyzing the existing labeling process in that specific case, we design a revised to-be process and develop a conceptual model from a value co-creation perspective. Finally, a functional prototype as an instantiation in the ITSM domain is implemented and evaluated through accuracy metrics and user evaluation. The results show that the new process increases the perceived value of both labeling quality and the perceived prediction quality. Thus, we contribute a conceptual model that supports the systematic design of efficient and interactive labeling processes in diverse applications of reinforcement learning systems.