Etsy Click-Through Rate Prediction
In this paper, 3 employees from Etsy analyze algorithms to predict click-through rates of ads in the Etsy website. Ads are displayed in a manner similar to that in most online merchant sites; a seller provides a budget, and the Etsy ad platform computes the bidding price for the seller’s given listing. Based on this bidding price, the listings are ordered and the sellers are charged according to a generalized second price auction. The seller’s end price depends on a cost-per-click model, so having a good model for the click-through-rate is essential for an effective ad management platform.
Etsy uses a data collection/training instance creation to model training to CTR prediction pipeline. Data collection from historical ad clicks is used to label listings with positive/negative labels, and the ads themselves (images of a product) are analyzed using Torch. A large part of the proposed system involves integrating many metrics to categorize and grade click through ads. This is interesting and relevant to the course material; the bidding system is almost exactly the GSP we learnt in class, and it exposes the different approaches to calculate the CTR that was assumed during the calculations from class.