Abstract The ABC classification technique is widely used to streamline inventory systems composed of thousands of stock-keeping-units (SKUs). A large number of methods belonging to different approaches have been developed to address the multi-criteria inventory classification (MCIC) problem. Two approaches are often discussed in the literature: the mathematical programming (MP) approach and the machine learning (ML) one. However, most of the developed research has only focused on providing classification methods to rank SKUs in an inventory system with limited interest in the original and most important goal of this exercise, which is achieving a combined cost-service inventory performance. Moreover, for the ML approach, most of the research has looked at supervised ML methods and very few studies have considered unsupervised ML methods, despite their potential to improve the cost-service performance of inventory systems without prespecifying the classes’ distribution. This paper analyses the empirical inventory performance of three unsupervised machine learning methods: K-means, agglomerative hierarchical clustering and gaussian mixture model. The performance of these methods is compared to that of well-known MP methods. We also put forward a new hybrid MP classification method with the criteria weights being set using a triangular distribution estimation method. The inventory performance of the proposed methods is empirically examined on two datasets containing more than 9000 SKUs. The results reveal that the unsupervised ML methods lead to a high inventory efficiency. We also provide empirical evidence on the benefit of a prior setting of the criteria weights in the MP method.
Zowid et al. (Sun,) studied this question.