Semi-automated estimation of weighted rates for e-commerce catalog quality monitoring
Product catalogs represent the backbone of e-commerce websites. Given these catalogs' constant evolution, we need to closely monitor the quality of their product information. Identifying defective product information, however, often requires human auditing, which makes catalog monitoring expensive. In this article, we investigate approaches for tracking weighted rates over time, here defined as the fraction of customer attention that goes to items with a particular defect. We focus on these metrics, given that to improve customer trust we need to minimize their exposure to listings with defective information. We assume that the gold standard for detecting defects comes from human auditors, but to avoid collecting audits at each point in time, we leverage existing machine learning classifiers. However, simply replacing human auditor decisions with automated predictions generally leads to large biases in the estimated weighted rates. We instead leverage classifiers while obtaining approximately unbiased and low variance estimators of the weighted rate of interest. We rely on being able to evaluate the quality of the classifier using audits at a baseline time, and then extrapolate its performance to the target times. We perform extensive simulation studies to stress-test our proposed estimation approaches under a variety of scenarios representative of our use cases. Our proposed estimation approach is related to the task of quantification in machine learning, and so we draw connections throughout the document.