Abstract
E-commerce marketplaces make use of a number of marketing channels likeemails, push notifications, etc. to reach their users and stimulate purchases.Personalized emails especially are a popular touch point for marketers toinform users of latest items in stock, especially for those who stoppedvisiting the marketplace. Such emails contain personalized recommendationstailored to each user's interests, enticing users to buy relevant items. Acommon limitation of these emails is that the primary entry point, the title ofthe email, tends to follow fixed templates, failing to inspire enough interestin the contents. In this work, we explore the potential of large languagemodels (LLMs) for generating thematic titles that reflect the personalizedcontent of the emails. We perform offline simulations and conduct onlineexperiments on the order of millions of users, finding our techniques useful inimproving the engagement between customers and our emails. We highlight keyfindings and learnings as we productionize the safe and automated generation ofemail titles for millions of users.