Tag: algorithm

The Art and Science Behind Spotify Wrapped

December 4, 2024

Spotify Wrapped is officially here, and with it comes your highly personal look back at how you soundtracked your year. 

We know your year in review evokes all the feels, but perhaps you’d like a bit more insight into all the ways your listening and our lists come together. From the skilled team of music editors who help develop Wrapped to the ace engineers who are evolving our tools to better fit your listening habits, it’s truly art and science working side by side. 

We caught up with JJ Italiano, Spotify’s Head of Global Music Curation and Discovery, and Molly Holder, Spotify’s Senior Director of Personalization, to reveal more about the magic behind Wrapped.

Happy Wrapped launch to you both! First, how does our annual experience highlight the work of both our brilliant tech team and the expertise of our editorial team? 

JJ: The work of Spotify’s editorial experts is grounded in a deep understanding of music culture. We are always working as a global team of music editors to identify and amplify trending songs, stories, and artists to the user—and, crucially, to contextualize those moments in curated playlists or other editorial offerings to help users form a deep, lasting connection to the track or artist. 

Spotify Wrapped gives listeners a chance to look back at the artist and track connections they’ve made over the course of the year through their personalized Wrapped list while also exploring some of the songs and artists that helped define the year in music culture more broadly through our editorial end-of-year Wrapped playlists.

Molly: The beauty of Wrapped is how much it reflects each individual user. Whether it’s your top song, total minutes listened, or favorite artist, 2024 Wrapped celebrates how you listened this year. 

We believe that great personalization fuels discovery: We’re powering nearly 2 billion music discoveries every day, connecting listeners with artists and genres they might never have found otherwise.

In addition to the work JJ and his team are doing, we’re excited to launch several brand-new personalized Wrapped features this year. From innovations like Your Music Evolution to the expansion of our AI-driven tools (such as AI DJ, AI Playlist, and Wrapped AI Podcast), users will have even more ways to explore the songs that defined their year.

JJ, can you share more about how the work of Spotify’s editorial team influences what users discover throughout the year, culminating in what shows up in their Wrapped? 

JJ: Over the course of the year, our editors around the world are constantly monitoring new music and cultural trends to identify important rising tracks and artists. We look at a vast range of qualitative and quantitative signals both on- and off-platform which inform our editorial decisions. 

We also continuously monitor in-playlist performance to understand what tracks and artists users are responding to. This helps signal what tracks should be added to more playlists. Your Wrapped may be comprised of artists that you’ve discovered anywhere on Spotify, including our editorial surfaces.

Molly, building on that, how do Spotify’s algorithms use those editorial selections to enhance personalized recommendations outside of Wrapped to ensure our users are discovering new music all year round? 

Molly: Spotify personalization is a mix of giving you what you like while pushing you a bit outside your comfort zone, too. We learn about what you like based on the ways you interact with Spotify, and our personalization technology takes into account a number of signals that you, as users, provide. For example, as you add songs to a playlist, listen to an entire song, skip a song, or engage with an artist, it sends us clear signals that help us tailor our programming to your taste.

But it’s also our job to determine when we should introduce you to something new. That’s when we consider certain additional factors, such as signs of popularity and how other users are engaging with tracks. We also rely on our editorial team, who play a crucial role in curating playlists and, through their knowledge, intuition, and expertise, help teach our machines. 

Recommendations are shaped by a number of signals, each contributing to the decision of whether a track should be suggested to a user. While these signals are weighted differently, listener satisfaction is always a priority, and we only recommend music we think listeners will want to hear. For more information on the most significant inputs that drive our personalization, you can learn more here.

How does Spotify use both human curation and algorithms to combat repetitiveness in recommendations?

Molly: Finding the right balance between familiarity and discovery is incredibly difficult. One reason is that music is different from other formats. On other platforms, you may consume a piece of content once and then never revisit it again. On Spotify, you can listen to the same song dozens of times, so we know that some repetitiveness is good.

But we don’t just paint by numbers. We listen to our users and the descriptive way they are saying they’d like to see us better balance familiarity with discovery, new with old, music with podcasts and audiobooks. We take that qualitative feedback to heart just as much as we weigh the quantitative signals when we make decisions about what to tweak in our algorithms. 

Can you share how human expertise and algorithms have worked together to propel a song or artist to success on Spotify?

JJ: One of this year’s most compelling examples would be Charli xcx’s BRAT album and the associated cultural moment, “brat summer.” Our editors were able to hear the album early and identified it as one of the top stories of the summer to come. As the album rollout and cultural conversation accelerated and branched off into new tracks and trends, our editors worked to reflect each moment rapidly and participate in the conversation holistically throughout our editorial surfaces.

BRAT not fitting neatly into any one genre gave us the opportunity to stretch our genre remits and work collaboratively to support her across multiple playlists and destinations on platform. Not only did Charli land on the cover of our New Music Friday playlist on album release day, but also on the covers of hot girl walk, Party Hits, and hyperpop, which helped us target the music to audiences outside of her traditional genre lanes. Watching the data allowed us to find more places for BRAT to be discovered on platform, and served to the right users.

Molly: As I mentioned, recommendations are powered by data. But there are instances—like when an artist releases a new track—where there’s little to no listening data available.

In those cases, we look at additional factors, such as signs of growing popularity. For example, if a song is climbing the charts, it becomes a valuable signal for our recommendation algorithms. 

Another key factor is input from JJ’s team, whose cultural expertise helps identify songs they believe our users will enjoy. And this year, we’ve seen some amazing breakout success from artists all over the world. Check out our 2024 Wrapped top lists to see what our listeners have deemed the songs, artists, and albums of the year! 

Explore your personalized 2024 Wrapped and learn more about this year’s campaign and experience on our 2024 Wrapped hub.

Responsibly Balancing What Goes Into Your Personalized Recommendations

March 6, 2023

Every month, tens of billions of discoveries happen on Spotify. Personalized recommendations play an important role in our ability to match listeners around the world with the right content, tracks, artists, or creators at the right moment. Behind the scenes, we combine human editorial expertise with a multitude of signals and systems with the aim of providing every listener with a unique and safe experience. 

At Spotify we focus on delivering recommendations that are relevant, encourage diversity in listening, and provide the opportunities for artist and creator discovery. We spoke with Henriette Cramer, Director of Algorithmic Impact, and Amar Ashar, Head of Algorithmic Policy—both members of the Trust & Safety team—for a deeper look at algorithmic impact and safety. 

Why focus on algorithmic impact? 

Henriette: Algorithmically programmed experiences like Discover Weekly, Release Radar, and Made for You Mixes, or even Search, provide opportunities for artists and podcast creators to grow their fan bases. But while machine learning and algorithms enable these really important opportunities, we know we have a responsibility to mitigate unintended harms, ensure we represent a very wide range of global creators on our platform, and understand our impact.

Understanding our algorithmic impact requires extensive internal and external collaboration, and we approach this space through three channels: research, product engagement, and collaboration with external partners. It’s an ever-evolving field, and we’re proactively working with Spotify teams and external stakeholders to continuously improve our approach as we continue to learn. 

What makes Spotify unique, from an algorithmic perspective? 

Amar: People often talk about the “Spotify Algorithm,” but that’s an oversimplification. In fact, Spotify’s personalization is a combination of a variety of algorithms, along with editorial and data curation teams, all contributing to a unique experience for each listener.

Spotify editors play a crucial role within this space by using their expert judgment to curate playlists and help artists find new fans. They also work with algorithms to create highly situational and personalized experiences. We call this “algotorial”—bringing both the editorial and algorithmic worlds together. This collaboration is critical to the Spotify experience. Think of it this way: Algorithms don’t go out to concerts, people do, which is why human expertise is an essential ingredient in our recommendations. 

We just released a new AI DJ that delivers a curated lineup of music alongside commentary around the tracks and artists. How are teams at Spotify working together to make sure the safety of recommendations is prioritized?

Henriette: In general, ensuring we approach Spotify recommendations responsibly requires close coordination between lots of teams across product, policy, legal, and research. We work with each of them to provide guidance that’s reflective of our algorithmic equity and safety goals, and we use various tools, such as algorithmic assessments, that help us identify and solve problems before they happen. 

Spotify’s DJ takes a unique approach by combining Spotify’s personalization technology, generative AI in the hands of music editors, and voice technology. The expertise of our editors is something that’s really important to our philosophy. As we launch new features, we aim for appropriate safety measures and processes to be in place. The product has been tested in a closed environment for a while, and now that we have launched this product as a beta, we’ll continue to study and improve the experience. 

How does your team work with external partners to improve Spotify’s personalized experience?

Amar: Engaging with research communities outside of Spotify is imperative to do our work. That’s why we also continue to share our findings with the wider community, collaborate across sectors, and ensure, as an industry, that we keep learning and evolving existing practices. 

We also work closely with external partners through Spotify’s Safety Advisory Council, which includes an interdisciplinary group of experts who advise us on safety topics and bring expertise on recommendations, responsibility, and safety from a global perspective.

What’s your go-to playlist?  

Amar: Discover Weekly, not only because it’s consistently a great playlist that has introduced me to new artists and genres, but also because I’ve been lucky enough to have worked with the team that’s built this flagship product.   

Henriette: So many! I love editorial playlists like Techno Bunker, Queens of the Blues, or New Orleans Brass to really get into a genre. Since I worked on voice projects in the past, it’s been really nice to play with the new DJ beta and see editorial, tech, and design work shine together as we continue to study how we can use new techniques responsibly.

Amplifying Artist Input in Your Personalized Recommendations

November 2, 2020

Listeners enjoy Spotify because we introduce them to music to fall in love with—including music they might not have found otherwise. In fact, Spotify drives 16 billion artist discoveries every month, meaning 16 billion times a month, fans listen to an artist they have never heard before on Spotify. We’re proud of that and are actively refining our algorithms to enable even more fan discoveries of new artists each month.

We’re able to make great personalized recommendations because of complex, dynamic systems that consider a wide variety of inputs about what you like—which we refer to as signals—and balance those signals in many possible different pathways to produce an output: the perfect song for the moment, just for you. 

This might sound complicated—and it is! Our personalized recommendations take into account thousands of types of signals: what you’re listening to and when, which songs you’re adding to your playlists, the listening habits of people who have similar tastes, and much more. In order to create algorithms that truly deliver the right song for the right time, we’re also taking into account less obvious factors: things like time of day, or the order in which you’re listening to songs or podcasts, or the release date of a song. 

Artists tell us they want more opportunities to connect with new listeners, and we believe our recommendations should also be informed by artists—their priorities and what they have to say about their music. And soon, we will roll out a test of a service that gives artists a say in how their music is discovered. 

In this new experiment, artists and labels can identify music that’s a priority for them, and our system will add that signal to the algorithm that determines personalized listening sessions. This allows our algorithms to account for what’s important to the artist—perhaps a song they’re particularly excited about, an album anniversary they’re celebrating, a viral cultural moment they’re experiencing, or other factors they care about. 

To ensure the tool is accessible to artists at any stage of their careers, it won’t require any upfront budget. Instead, labels or rights holders agree to be paid a promotional recording royalty rate for streams in personalized listening sessions where we provided this service. If the songs resonate with listeners, we’ll keep trying them in similar sessions. If the songs don’t perform well, they’ll quickly be pulled back. Listener satisfaction is our priority—we won’t guarantee placement to labels or artists, and we only ever recommend music we think listeners will want to hear. 

We’re testing this to make sure it’s a great experience for both listeners and artists. To start, we’ll focus on applying this service to our Radio and Autoplay formats, where we know listeners are looking to discover new music. As we learn from this experiment, we’ll carefully test expanding to other personalized areas of Spotify. 

We believe this new service will unlock more discoveries than ever before. Our recommendations rely on signals from you, so keep on listening to what you love!