Toward Automated Detection of Economically Significant Anomalies in Digital Storefronts: Evaluation Criteria and Conceptual System Architecture

E-commerce has become an integral part of our lives, with online stores and marketplaces offering products through Digital Storefronts. The Digital Storefront and user behavior on it form a dynamic system, subject to constant changes that ultimately impact the storefront’s economic performance, such as the expected average purchase value or conversion rate. Changes in product assortment, technical issues, and many other factors can cause deviations in performance indicators. In cases of critical deviations, companies are forced to involve analysts to identify the root causes, which can take considerable time and lead to financial losses due to delayed reactions. This paper addresses the problem of detecting and localizing economically significant anomalies in Digital Storefronts based on user behavior analysis. It reviews the historical development of the field through the lens of Web Mining and identifies key data structures used for modeling behavior. Particular attention is given to evaluation criteria for anomaly detection methods—such as adaptability, interpretability, automation, and sequential awareness—as well as to the limitations of existing method classes. As no single method class meets all criteria, the proposed approach relies on a modular architecture that decomposes the task into specialized subsystems. The paper outlines the conceptual architecture of such a system, integrating time series analysis (e.g., ARIMA, HMM), unsupervised learning, graph-based models (e.g., GCN), and deep learning techniques (e.g., LSTM, Transformers), with a focus on actionable and interpretable outputs for decision support. © 2025, International Consortium for Electronic Business. All rights reserved.