Abstract
Online marketplaces execute large volume of price updates that are initiatedby individual marketplace sellers each day on the platform. This pricedemocratization comes with increasing challenges with data quality. Lack ofcentralized guardrails that are available for a traditional online retailercauses a higher likelihood for inaccurate prices to get published on thewebsite, leading to poor customer experience and potential for revenue loss. Wepresent MoatPlus (Masked Optimal Anchors using Trees, Proximity-based Labelingand Unsupervised Statistical-features), a scalable price anomaly detectionframework for a growing marketplace platform. The goal is to leverage proximityand historical price trends from unsupervised statistical features to generatean upper price bound. We build an ensemble of models to detect irregularitiesin price-based features, exclude irregular features and use optimized weightingscheme to build a reliable price bound in real-time pricing pipeline. Weobserved that our approach improves precise anchor coverage by up to 46.6% inhigh-vulnerability item subsets