Enhancing Out-of-Distribution Detection for Retail Time-Series Data Using Entropic Methods

Machine learning models are typically developed under the “closed-world” assumption, where training and testing data originate from a consistent distribution. However, in real-world scenarios, especially in the retail domain, this assumption can become problematic due to the frequent introduction of new products, seasonal promotions, and irregular sales events. When models encounter out-of-distribution data inputs, predictions can become overly confident or entirely incorrect. While existing out-of-distribution detection methods primarily focus on image-based datasets, challenges associated with numerical, high-dimensional, and heterogeneous retail time-series data remain largely unexplored. To address this gap, this study proposes an enhanced Entropic Out-of-Distribution Detection framework tailored specifically for dynamic retail environments. By trans-forming time-series sales data into spectrogram representations and leveraging the IsoMax+ loss function, our approach im-proves uncertainty calibration and robustness without requiring labeled out-of-distribution data or additional post-hoc calibration techniques. Experimental results, conducted on a large-scale retail dataset from Vietnam, demonstrate that the proposed Entropic Out-of-distribution detection framework significantly outperforms traditional out-of-distribution detection methods in terms of detection accuracy and inference efficiency, providing a scalable and practical solution for real-time retail applications. Our approach achieves strong performance with an F1-score of 88% and an AUC of 91%, highlighting its promising applicability across diverse business scenarios.