Warming Up to Cold Start Personalization

Abstract

Smart agents face abandonment if they are unable to provide value to the users from the very first interaction. Existing smart agents take time to learn about new users before they can offer them personalized services. We present a method for learning personalization information about users quickly and without placing unnecessary hardship on them. Our method enables smart agents to pick which questions to ask the user when they first interact to maximize the agent's overall knowledge about the user. We demonstrate our method on two publically available US census datasets containing 172 user variables from 1,799,394 training and 1,618,489 testing users. The questions selected using our method improve the agent's accuracy when inferring information about future users, including information that they did not ask about. Our work enables smart agents that assist the user with personalized services soon after they start interacting.