Accelerating Creator Audience Building through Centralized Exploration

February 23, 2024 Published by Olayinka S. Folorunso, Maksym Lefarov, Buket Baran, Guilherme Dinis Jr., Tonia A. Danylenko, Gosta Forsum, Yu Zhao, Lucas Maystre

TL;DR: Across a broad spectrum of product features, Spotify’s array of recommender systems play a pivotal role in tailoring personalized user experiences and helping creators grow. These systems undertake local exploration to enable fresh content on the platform and provide a unique pathway for audience growth. At the RecSys 2023 conference, we presented work aimed at enhancing the efficiency of this exploration process, which is now in production. We developed a new workflow by centralizing content exploration and then propagating the exploration learning to existing, decentralized recommender systems. We take a creator-centric perspective, and online experiments show that this approach can significantly reduce the time it takes for new content to reach its full potential and expedite audience growth for creators.

Audience building

On today’s digital content platforms, the large number of creators and the diversity of content can make it challenging for creators to find their audience and for listeners to discover their next favorite listen in a timely manner. At Spotify, we aim to unlock the potential of human creativity by giving a million creative artists the opportunity to live off their art and billions of fans the opportunity to enjoy and be inspired by these creators. To achieve this vision, we are helping listeners find new content, as well as assisting creators in finding an audience quickly when their content is published on the platform. The latter requires exploring among potential listeners to understand how to help the content gain traction and, ultimately, increase engagement. This is particularly important for emerging creators, who might have a high potential but no established audience yet. 

At a high level, the underlying technical challenge of content exploration is the well-documented cold-start problem. When new content is published on the platform, and before observing a critical amount of user feedback, it is challenging for recommender systems, typically trained on user-content interaction data, to start surfacing the content to the right audience in a timely manner. All Spotify recommendation systems have been designed with mitigating cold-start problems, either directly or indirectly, through utilizing a wide range of algorithms and strategies to discover and promote new creators and their content.

Decentralized recommendation ecosystem

Spotify aims to deliver a coherent experience to the end user, but underneath, this experience combines many different recommender systems that power different interaction modes and fulfill different product goals. Organizationally, these different systems are owned by distinct teams, as this gives each team the necessary focus to create a great experience for each interaction mode. 

To use an example from Spotify’s recommendations, there is one recommender system that powers the Discover Weekly playlist and another one that powers the Daily Mix playlists. These systems may have multiple different optimization objectives; for example, Discover Weekly is optimizing for discovery—introducing tracks, artists, and genres that are novel to the user—whereas the Daily Mix personalized playlists are based on the user’s favorite content with some new music they may enjoy.

Each system locally makes efforts to ensure the diversity of recommended content and timely discovery of new creators. The local learnings are integrated and shared implicitly through organization-wide foundation models that utilize data from the whole organization. Continuing on the example from above, if Discover Weekly successfully connects a Spotify user to a new creator, this connection could potentially be picked up and subsequently propagated to that user’s Daily Mix playlists. This would further consolidate the connection between the user and the creator.

However, with the growing number of creators and content on the platform, we recognised a growing need to more explicitly foster audience building in a more centralized way.

Centralized exploration

To improve the audience-building at Spotify, we borrowed a few concepts from the Reinforcement Learning (RL) domain. Explore-exploit – as it is known in the industry – is one of the key ideas in RL that captures the fundamental trade-off that an agent faces when making decisions in uncertain situations: balancing the need to learn about the outcomes of making certain decisions (exploration) while simultaneously maximizing the returns based on the existing knowledge (exploitation). Explore-exploit provides a natural framework for addressing the cold-start problem in recommender systems. 

Most existing applications of explore-exploit to recommender systems assume exploration and exploitation happen in a standalone recommender system controlled by a single decision-making policy. As already alluded above, this is not applicable to Spotify’s setting, where a large number of distinct recommender systems operate in a distributed way with different objectives. We therefore decided to split explore-exploit into a pure exploration phase to estimate a potential audience of new content from aspiring creators and a follow-up exploitation phase to build this audience by acting on discovered information. Furthermore, we centralized the exploration in the recommendation ecosystem and and then relayed the information learned during the exploration phase to the distributed exploitation components, consisting of all the recommendation sub-systems on Spotify.

In theory, a centralized exploration component should allow for a more efficient and effective allocation of exploration impressions by coordinating the learning across Spotify surfaces.  There are two main challenges: 1) how to allocate exposure for new content that appears on the platform while respecting a given “budget” of impressions; and 2) given the diversity of Spotify’s recommender ecosystem, how to propagate the learned information to the exploitation components. To account for these and allowing further developments, we generated different representations of learning that are propagated for use in the downstream exploitation systems:

• Raw exploration signals: the full set of impressions and corresponding user interactions collected during exploration.
• New content representations: fixed-size vector features that can be used as inputs to predictive models.
• Personalized recommendation candidates: many recommender systems use a two-stage setup with the first stage, retrieval, generating a small set of relevant candidates and the second stage ranking the candidates into a final ordered list. We also generated a personalized candidate pool, specific to new content, that is based on learnings from exploration and that can be easily combined with existing retrieval models.

After the centralized exploration phase ends, the generated signals and knowledge on the new content flow to the distributed recommender sub-systems for exploitation (i.e. leveraging the learnings to optimize the recommendation of explored content). Equipped with this knowledge, these subsystems can make sense of and give adequate consideration to newly published content in a quick and timely manner. At the same time, these systems can still operate based on their own specific knowledge, goals, and capabilities, and they can independently and autonomously make recommendation decisions aligned with their distinctive performance goals. 

The centralized exploration system has been deployed in production at Spotify, and the learnings are integrated into some distributed recommendation systems for exploitation. For instance, in an online A/B test, we observed an increase in the number of listeners by a factor of 10 on the explored content without negatively impacting the local metrics relevant to the specific system where the exploitation occurred—this is an exciting step on our journey to help more creators grow their audience.

What’s next?

We continue to evolve.  We are looking at evaluation criteria for centralized exploration, as well as ensuring that the recommender systems are reactive and sensitive enough to the signals and learnings generated by exploration. We will continue our effort to make Spotify a place where up-and-coming creators can build their audience in an effective and efficient manner.

For more information, please check our paper:
Accelerating Creator Audience Building through Centralized Exploration
Buket Baran, Guilherme Dinis Junior, Tonia Antonina Danylenko, Olayinka S. Folorunso, Gösta Forsum, Maksym Lefarov, Lucas Maystre, Yu Zhao
RecSys 2023