A Machine Learning and Multi-Criteria Decision-Making Approach to Cycle Counting

Background: Inventory record inaccuracy (IRI) causes discrepancies between physical and digital inventories, leading to production delays and customer dissatisfaction. Cycle counting, in this context, is a common corrective action. Pareto-based ABC analysis is widely used to decide which items to inspect, but it often oversimplifies inventory decisions, and recent studies suggest that multi-criteria decision-making (MCDM) and machine learning (ML) may offer more effective solutions. Methods: This study applies the analytic hierarchy process (AHP) method, combined with K-means (AHP-K), to classify stock-keeping units (SKUs) into three groups with distinct counting policies. A selection procedure is then applied to identify an optimal ML algorithm and compare its classification with the original AHP-K results; each model in this phase is trained on a subsets of 100 SKUs. A Veto method is also introduced to improve output consistency for both AHP-K and the best ML method, and a comparative cost evaluation is presented. Results: The ML-AHP-K-Veto classification achieves over 90% accuracy. Analysis of a dataset of 12,863 SKUs from a mechanical manufacturing company shows minimal cost differences between ML-based and MCDM classifications, but significant differences compared to Pareto-based costs. Conclusions: ML can effectively address IRI, supporting the development of pure ML applications, including decision-maker (DM) preferences, to manage cycle counting strategies.