Improving Retail Sales through Unsupervised Collective-Contextual Anomaly Detection: A Deep Reconstruction Autoencoder for Network-Wide Sales Analysis
Abstract
Vast sales data in the retail sector present opportunities to optimize operations and increase revenue. Consequently, studies have leveraged these data to predict sales, detect anomalies, segment customers, and similar. Anomaly detection techniques have been proposed to identify deviations in sales patterns at both the product and product-store levels. However, these studies did not consider sales patterns across similar stores, which could greatly benefit the retailers. For instance, if a particular store outperforms similar stores in sales of a specific product, best practices could be established and shared throughout the retail network, ultimately increasing overall sales. Consequently, this thesis proposes an unsupervised collective-contextual anomaly detection framework for the identification of outperforming stores in massive sales data from a large retail network. The proposed approach addresses the gap by decoupling the problem into two components. Collective anomalies are detected by leveraging autoencoders to identify patterns that deviate from normal behavior. Then, contextual anomaly detection leverages the designed similarity metric, which operates at the product level, to identify similar stores and examine patterns among them. Experiments using a dataset from a major Canadian retailer demonstrate the success of the proposed approach, achieving 90\% overall precision.