EXPRESS: From Chit-Chat to Conversion: Cost-Efficient Conversational Recommendation via Retrospective Learning and Isolated Reflection

Conversational recommendation systems help digital platforms elicit user preferences through multi-turn interaction, but they often face a practical tension between recommendation accuracy and interaction cost when users provide vague or low-information inputs. This study follows a design science approach to develop a cost-conscious conversational recommendation artifact for such ambiguous service encounters. The artifact is guided by two design principles. First, retrospective outcome-guided supervision uses successful historical conversion trajectories as downstream anchors for improving responses at earlier low-information states. Second, isolated reflection separates recommendation generation from consistency checking, enabling diagnostic regeneration when candidate responses conflict with previously stated user constraints. In implementation, the artifact instantiates these principles through parameter-efficient adaptation of open-source large language models and prompt-isolated role switching over a shared backbone model. We validate the artifact on three conversational recommendation datasets covering movie and e-commerce settings, using outcome efficacy, interaction efficiency, process analysis, ablation analysis, robustness checks, and human evaluation. Here, interaction efficiency refers to user-visible dialogue turns, while computational burden is reported separately through training, memory, and latency measurements. The empirical results show that the artifact improves recommendation effectiveness while maintaining a favorable balance between interaction burden and conversion outcomes under the tested conditions. Process-level analysis further characterizes how recommendation behavior evolves over the course of dialogue. This study contributes transferable design principles for embedding retrospective preference optimization and isolated reflection into conversational recommendation systems, offering operational guidance for platforms seeking to improve personalization under constraints on interaction cost.