Application of Product Recommendation Systems in the B2B Distribution Industry*
Jeong Soo Kim, Jeongsoo Kim, Myounggyun Jang, Juyoung Kim, Juyoung Kim · Journal of Channel and Retailing · 2025
Purpose: In the B2C sector, recommendation systems have been widely adopted to enhance customer decision-making and purchasing experiences by providing personalized product suggestions based on user preferences. However, the implementation of such systems in the B2B domain remains limited due to challenges such as low transaction frequency, data sparsity, and heterogeneity in pricing and contractual conditions among buyers. This study seeks to empirically assess the applicability of recommendation systems in the B2B distribution industry and compare alternative collaborative filtering algorithms to identify an effective modeling approach tailored for B2B contexts. Research design, data, and methodology: Using a dataset comprising three years of transaction records from a wholesale distribution center serving small and medium-sized retailers, this study develops and evaluates collaborative filtering-based recommendation systems. Three algorithmic approaches, similarity-based collaborative filtering, full singular value decomposition (Full SVD), and truncated singular value decomposition (Truncated SVD) are compared. Additionally, retailer segmentation is performed through K-means clustering based on purchase frequency and transaction volume. To address the cold-start issue, a hybrid model integrating collaborative and content-based filtering is proposed. Eight models are constructed to account for both algorithm type and clustering presence, and their performance is evaluated using a training-validation framework. Results: Empirical findings indicate that the truncated SVD model yields the highest accuracy in product recommendations by effectively addressing the issue of data sparsity and enabling new product discovery. In contrast, similarity-based collaborative filtering demonstrates superior performance in predicting purchase quantities but exhibits a tendency to recommend previously purchased items, thereby limiting recommendation diversity. Clustering retailers by purchasing behavior enhances the precision of quantity prediction and enables more personalized recommendations, though at the expense of product diversity. The hybrid approach mitigates the cold-start problem by leveraging product attribute data, resulting in improved overall recommendation accuracy. Nevertheless, the predicted purchase quantities were consistently lower than the actual quantities observed. This discrepancy likely stems from the methodological adaptation of product recommendation techniques for quantity prediction, highlighting the need for dedicated models specifically designed for accurate quantity forecasting. Conclusions: This study provides empirical validation for the adoption of recommendation systems in the B2B distribution sector. Among the evaluated approaches, singular value decomposition-based collaborative filtering is particularly effective in mitigating data sparsity and generating novel product recommendations. While clustering enhances personalization and quantity prediction accuracy, it also introduces trade-offs in recommendation diversity, necessitating strategic deployment depending on organizational goals. The study contributes to the advancement of data-driven decision-making in B2B e-commerce and supply chain management. Future research should explore the development of specialized quantity prediction frameworks and the integration of advanced techniques such as deep learning and reinforcement learning to further enhance the sophistication and applicability of B2B recommendation systems across diverse industrial settings.