Learning to target advertisements at Spotify

Abstract: The goal of this thesis work is to develop a methodology to optimize advertisement targeting inside the Spotify platform. By understanding the relevance of advertisements to users, the advertisement efficacy can be improved leading to an overall reduced advertisement load, to a better user experience, and to an increased advertisement revenue. Three types of advertisement are delivered inside the Spotify platform: house advertisements, which promote the Spotify premium subscription, label advertisements, which promote artists and albums and commercial advertisements, which promote external brands or products. For each type, machine learning models were implemented to optimize the matching of advertisements to users. The house advertisement targeting focused on maximizing the conversions of users from free to premium subscriptions. Unlike traditional advertisement targeting models, which focus on estimating the user probability of converting, the proposed approach focuses on estimating the change in probability of converting when an advertisement is delivered, so that only impressions producing a true uplift are delivered. The targeting of label advertisements optimized the number of users starting to listen to the advertised music. The implemented approach delivers relevant advertisements based on an evaluation of the affinity between the user's music and the advertised music and on the user's listening behavior. The commercial advertisement targeting optimized the number of clicks on advertisements. Given that commercial advertisements advertise external products, the main challenge faced was the lack of relevant data to inform the targeting. The implemented approach tries to deal with this problem by combining feature based methods with collaborative filtering methods. The main contribution of this thesis work is the implementation of machine learning models to improve advertisement targeting inside the Spotify platform. In particular, the proposed methodology uses uplift modeling, with a modified approach to handle bias in the training data, and also makes use of meta-data to better understand the context of the campaigns.