08: Humans in the Loop

Get ready: we’re diving into machine learning. Hear how we’re improving personalization with reinforcement learning (RL), what makes ML engineering so different from other kinds of software engineering, and why machine learning at Spotify is really about humans on one side of an algorithm trying to better understand the humans on the other side of it.

Spotify’s director of research, Mounia Lalmas-Roelleke, talks with host Dave Zolotusky about how we’re using RL to optimize recommendations for future rewards, how listening to more diverse content relates to long-term satisfaction, how to teach machines about the difference between p-funk and g-funk, and the upsides of taking the stairs.

Then Dave goes deep into the everyday life of an ML engineer. He talks with senior staff engineer Joe Cauteruccio about what it takes to turn ML theory into code, the value of T-shapedness, the difference between inference errors and bugs, using proxy targets and developing your ML intuition, and why in machine learning something’s probably wrong if everything looks right.

Plus, an ML glossary: our guests educate us on the definitions for cold starts, bandits, and more. This episode is the first in a series about machine learning and personalization at Spotify.

Learn more about ML and personalization:

• Listen: Spotify: A Product Story, Ep.04: “Human vs Machine”
• Watch: TransformX 2021: “Creating Personalized Listening Experiences with Spotify”

Recent publications from Spotify Research:

• “Variational User Modeling with Slow and Fast Features” (Feb. 2022)
• “Algorithmic Balancing of Familiarity, Similarity, & Discovery in Music Recommendations” (Nov. 2021)
• “Leveraging Semantic Information to Facilitate the Discovery of Underserved Podcasts” (Nov. 2021)
• “Shifting Consumption towards Diverse Content on Music Streaming Platforms” (Mar. 2021)

Read what else we’re nerding out about on the Spotify Engineering Blog: engineering.atspotify.com

You should follow us on Twitter @SpotifyEng and on LinkedIn!