Summary
Millions of listeners around the world use Spotify to listen to audio experiences that are personalized to their taste, yet relevant in the broader cultural context. Spotify is a unique product because unlike many other eyes the first app, a majority of Spotify experience is phenomenological – it happens in listeners’ ears, minds and bodies, after they hit “Play”. Why do people form Meaningful Connections™ with some type of audio content, but not all, and how do we ensure that our ML models respect the nuances of what makes listening to music and podcasts enjoyable? This talk will shed light on such topics and leave the audience with ideas, challenges and possible best practices for using Mixed Methods Research in the age of ML to build delightful product experiences.
Key Insights
-
•
Personalization at Spotify happens on multiple layers: content sequencing, merchandising, and overall experience.
-
•
Spotify’s playlists balance algorithmic recommendation and editorial curation for personalized yet communal music experiences.
-
•
User behavior signals like song skipping are commonly assumed negative but are contextually complex and can be positive.
-
•
Mixed methods research at Spotify integrates user research and data science, with an emerging role for machine learning as a third component.
-
•
Distinct personas like 'Nick' (music collector) and 'Shelley' (casual listener) exhibit different skipping behaviors and trust models.
-
•
Skipping meaning depends on user goals: active music exploration vs passive listening.
-
•
Contextual factors such as environment and social setting affect user interactions and should influence model signals.
-
•
Machine learning models need to incorporate complex, nuanced human behaviors, mental models, and trust dynamics.
-
•
Investing in the right raw data signals is a multi-year strategic decision with major impacts on both user experience and business outcomes.
-
•
User research teams at Spotify have grown significantly as the importance of understanding human factors in ML-driven products is recognized.
Notable Quotes
"Machine learning used to be back-end technology, but now machines decide a lot of consumer experiences."
"If machines are crafting the user experience, why do they need me as a researcher? That was my big question."
"In the old world, radio DJs curated music for everyone; now every individual gets a unique playlist."
"Skipping a song is probably the most benign behavior, but it’s also the most misunderstood signal."
"Nick says, no one else can make your perfect playlist; you have to make it yourself."
"Shelley says, I don’t know what I want, but I know I want something good and I want Spotify to do the work."
"Some signals can be measured, some cannot, and some should not be measured in machine learning."
"Human behavior is complex and machine learning models need to take into account nuances, biases, trust, and mental models."
"We are now using mental model diagrams and journey maps physically in machine learning teams."
"There hasn’t been a better time to be a user researcher in a machine learning-driven product world."
Or choose a question:
More Videos
"A 101 conversation over coffee yields better insights than written reviews in 360 assessments."
Adam Cutler Karen Pascoe Ian Swinson Susan WorthmanDiscussion
June 8, 2016
"Playing politics in UX leadership is about maximizing relationships ethically to advance your agenda, not about being underhanded."
Peter MerholzThe Trials and Tribulations of Directors of UX (Videoconference)
July 13, 2023
"Governance is about decision making, not workflow processes."
Lisa WelchmanCleaning Up Our Mess: Digital Governance for Designers
June 14, 2018
"Policy change is the backbone of effective climate strategies in urban areas."
Vincent BrathwaiteOpener: Past, Present, and Future—Closing the Racial Divide in Design Teams
October 22, 2020
"If you get a perfect score on your OKRs, it means you didn’t set your sights high enough."
Brenna FallonLearning Over Outcomes
October 24, 2019
"Never be afraid to get into good trouble; start by asking why."
Tricia WangSpatial Collapse: Designing for Emergent Culture
January 8, 2024
"Proto personas created by cross-department participants helped us build unbiased, relevant survey questions."
Edgar Anzaldua MorenoUsing Research to Determine Unique Value Proposition
March 11, 2021
"We manage the environment in which knowledge distribution takes place, not the process itself."
Designing Systems at Scale
November 7, 2018
"Decades worth of agricultural data had to be thrown out because they lacked control groups and statistical rigor."
Erin WeigelGet Your Whole Team Testing to Design for Impact
July 24, 2024
