ConfigRec: An efficient recommendatory configuration design method for customized products

Abstract

Product configuration design is essential in mass customization and enables the rapid selection of configurable components to assemble a desired product. However, existing configuration methods struggle to balance customization flexibility with production efficiency. The configurators process multiple components and orders sequentially, leading to extended computation times. Additionally, component coupling relationships introduce extra costs and complexity. To address these challenges, we propose ConfigRec, an end-to-end recommendatory configuration design method that leverages the parallel computing capabilities of deep learning. Specifically, our approach: (1) constructs specialized parameter embeddings for components by encoding diverse design parameters; (2) decouples complex relationships within the product configuration tree through top-down and bottom-up message passing, while capturing implicit dependencies using a linear attention mechanism; and (3) predicts instance recommendation scores and generates a customized Engineering Bill of Materials based on a formally defined configuration decision law. A real-world case study on elevator products demonstrates that ConfigRec achieves up to 99.51 % accuracy within seconds. The proposed method is interpretable, efficient, and highly accurate, significantly reducing customized product delivery times.