TastePaths: Enabling deeper exploration and understanding of personal preferences in recommender systems
March 18, 2022 Published by Savvas Petridis, Nediyana Daskalova, Sarah Mennicken, Samuel F Way, Paul Lamere, Jennifer Thom
Existing recommender systems are limited in the ability to help us grow and understand our personal music preferences
Recommender systems are ubiquitous and influence the information we consume daily by helping us navigate vast catalogs of information like music databases. However, they mostly take a linear approach of surfacing content in ranked lists (such as playlists), without showing any context such as how the artists are related or what the sub-genre of each track is. This limits their ability to help us grow and understand our personal preferences. What if we had a way to visualize that information and explore music more actively while sparking exciting discoveries? This is just an early exploratory idea, but it might have implications on how we at Spotify help people discover new music, or how we support the editors that create our special playlists.
To better understand how users explore and understand the music genres they listen to, and how we can better support them in their exploration processes, we developed TastePaths: an interactive web tool that helps users explore an overview of the genre-space via a graph of connected artists.
How can we best help music listeners explore a novel music space? Let’s ask the experts
To learn how it is best to explore a novel genre, we started our research by interviewing five music curators at Spotify who delve into new genres and figure out what is necessary to learn about a genre on a daily basis. This helped us identify three main findings that we turned into our design goals:
- In order to give users a meaningful starting point for exploring a genre, we should help anchor them with artists they already know and listen to.
- In order to contextualize the genre and help users understand it and its components, we should present an overview depicting the genre-space and its subgenres.
- In order to allow users to easily assess what parts of the genre they like, we should have a quick and convenient way to deep-dive into an artist’s work while being able to seamlessly go back to exploring.
Incorporating what we learned from the experts into the design and implementation of TastePaths
By leveraging Spotify’s Web API, as well as a custom version created by Paul Lamere, TastePaths generates a force-directed graph of 150 related artists and assists the user in exploring and making sense of it. TastePaths helps a user explore a genre by basing exploration to three of their most frequently listened to artists in that genre (“anchor artists”, which appear as black dots in the graph). To help users make sense of all the nodes, TastePaths then clusters the artists and presents a legend, displaying each cluster’s three most representative sub-genres. Users can listen to the artist’s tracks by hovering their mouse over the nodes, and they can also add them to a playlist.
What do real Spotify users think of TastePath’s visual interface?
In the second part of our research, we explored how real users interacted with the TastePaths system. We recruited 16 participants from the dscout (ages 19 to 53), from diverse backgrounds. To be eligible for the study, they had to have a Spotify premium account, be interested in exploring new music, and listen to discovery-focused playlists (Discover Weekly & Release Radar) at least once in the last three months.
Through our analysis, we identified four main themes: (1) personalization is key, (2) best discoveries are between or on the edge of genres, (3) users want more control: human-in-the-loop growing and pruning of the graph, (4) improved recommendation explainability through mental map.
Theme 1: Personalization is key
In regards to our research question about the role of personalization in open-exploration, we found that personalization is key. Participants were on average more interested in exploring the personalized version of TastePaths and felt they more easily discovered new artists. Finally, users wanted even more personalization. They wanted more of their past listening data reflected in the network. Multiple users expressed interest in a heat-map feature that would help them prioritize clusters to explore.
Theme 2: Best discoveries are between or on the edge of genres
We were also interested in what strategies users might employ when removing the linear constraints. We found that participants encountered their best discoveries in between or on the edge of genres. Artists in between two genres captured essences of two musical styles, which led to exciting discoveries. Meanwhile, artists on the edge of a cluster would often be lesser-known and contain interesting deep-cuts in a sub-genre.
Theme 3: Users want more control: human-in-the-loop growing and pruning of the graph
Users also expressed interest in more control, desiring to grow the network where they found music they enjoyed and prune portions they found less interesting. They also imagined an adaptive version of the guide, where the path would change as they provided feedback on artists.
Theme 4: Improved recommendation explainability through mental map
We found that using TastePaths helps users understand the variance of music within a genre, what they liked and disliked in a genre, as well as the vocabulary to describe their interests. This new knowledge helped them feel better equipped to understand where their recommendations were coming from. And beyond this new knowledge of sub-genres, users also wanted to learn more, including the history of the genre, its sonic characteristics, and influential artists.
What does this mean for future recommender systems?
Our study highlights a few important points for discussion. First of all, we see the need for more expressive and natural feedback from users so that they can tailor how the system represents and understands their interests. By enabling expressive feedback, we can better inform the algorithms powering popular recommendation systems. One finding from our user study was that particularly rich and interesting discoveries would lie either between two clusters or on the edge of a cluster. This information could be used as implicit feedback to generate discovery playlists to enable further exploration with minimal effort.
Additionally, while people used TastePaths to explore their interests more deeply, they also really liked that it was a finite list of nodes, which gave them a sense of closure (unlike an endless feed of media content). One way to further support user agency in their consumption could be to help them form goals on how far or how long to explore during the session to encourage growth and agency as opposed to longer listening sessions.
Overall, we created TastePaths as an interactive web tool that helps users deeply explore and understand the music genres they listen to. Future tools in this space can investigate how to better incorporate learning into exploratory search, how to incorporate more closure and goal-fulfillment in recommendation systems, and how to support users in modifying the system’s representation of their taste and interests.
We would like to thank all of the curators and music editors who helped us with their feedback and insights about the tool!
More details about this work can be found in our paper:
TastePaths: Enabling Deeper Exploration and Understanding of Personal Preferences in Recommender Systems
Savvas Petridis, Nediyana Daskalova, Sarah Mennicken, Samuel F Way, Paul Lamere, and Jennifer Thom