Explainable Anomaly Detection in Retail Perpetual Inventory Systems Using Shap-Enhanced Isolation Forests

The systems that rely on correct data can also be afflicted with anomalies related to crime, barcode ‎scan problems, user error, and process difficulties that, if not detected can lead to financial losses ‎and disruptions to supply chains. Traditional anomaly detection systems identify potentially ‎fraudulent records but have little accountability through transparency. As such, managers have little ‎trust in the "black-box" results generated. This study provides one way of applying current ability in ‎isolating the best anomalies - isolation forests, one of the leading anomaly detection strategies, and ‎SHAP (SHapley Additive exPlanations), one of the premier explainable AI. They demonstrate not ‎only that the isolation forest aims to estimate the likelihood that a record is anomalous, they make ‎clear the factors within the records that contribute to each violation - whether that is increased sales ‎from promotions, negative value inventory, or promoting unusual sequences of transactions. Our ‎datasets consist of real retail inventory records, which shows how well an isolation forest + SHAP ‎framework performs compared to other models well known to anomaly detection. By making these ‎models explainable, they not only become more effective, the new situates them under a decision ‎support system given their manageria1 trustworthiness to reduce shrinkage, and provide greater ‎operational resiliency to retail firms‎.