Enhancing Recommender Systems: A Practical Study on Loss Function Optimization in B2B Commerce

In this study, we present our work on developing a stock recommendation system for sales of industrial parts to customers at a mayor technical center in Jordan. The system leverages 40,000 transactions collected over a two-year period which are analyzed by machine learning to generate recommendations for internal decisions of whom to offer which product according to similarities between customers, items and past transactions. We compare collaborative filtering using matrix factorization with a hybrid model that exploits embedded item and user features such as categories and temporal buying behavior. Several loss functions are compared for both models, and we show that depending on the specific need of the user, either model shows better recommendation results. The study shows the importance of model selection and parameter fine-tuning when implementing a recommendation system in a B2B context, while also pointing out possible pitfalls in building such a system with real-world data.