TechPrufer
Get started
Spotify
Spotify logoSpotify · Feature trail

ML-driven Personalization Systems

9 posts · Dec 2014 – Mar 2023

What this trail covers

Spotify's ML-driven personalization systems continue to evolve, with a focus on leveraging reinforcement learning (RL) to optimize for long-term user satisfaction. This involves processing vast amounts of user data to build sophisticated models that predict user needs and preferences across various contexts, aiming to create a holistic and fulfilling audio experience. The system also considers factors like time of day, user activity, and device to tailor recommendations, ultimately benefiting both users and artists. The 'Algotorial' approach combines human editorial expertise with algorithmic recommendations to create personalized editorial playlists, balancing broad appeal with individual taste.

All Spotify trails
  1. Humans + Machines: A Look Behind the Playlists Powered by Spotify’s Algotorial TechnologyThis post introduces 'Algotorial' technology, which combines human editorial expertise with algorithmic personalization to create playlists. Editors define user needs and create 'pools' of relevant songs, leveraging their musical and cultural knowledge. Algorithms then select and order tracks from these pools for individual users. The system adapts based on user engagement (listening, skipping, saving) to refine taste profiles and recommendations. Personalization is highlighted as a key driver of user satisfaction, with 'Algotorial' playlists scaling editorial expertise for personalized experiences.
  2. How Spotify Uses ML to Create the Future of PersonalizationThis post introduces the application of reinforcement learning (RL) to Spotify's personalization systems. It details how RL models are used to predict and maximize long-term user satisfaction by guiding users towards a diverse and fulfilling content diet, rather than just short-term gratification. The system simulates various scenarios to train the RL model, aiming to create a more holistic audio experience that benefits both users and artists by increasing content discovery and engagement.
  3. The Rise (and Lessons Learned) of ML Models to Personalize Content on Home (Part I)This post details the evolution of Spotify's Home page recommendation ML models, specifically addressing the challenges of unifying training and serving data. It introduces the concept of a single transformation implementation for features to prevent discrepancies between training and serving data, exemplified by the Shortcuts Model migration. It also highlights the adoption of Tensorflow Data Validation (TFDV) for daily comparison of training and serving data schemas and feature distributions, with Chebyshev distance used for alerting on drift. The post emphasizes the shift from batch predictions to real-time serving and the automation of feature logging, data validation, and model retraining within Kubeflow pipelines.
  4. The Rise (and Lessons Learned) of ML Models to Personalize Content on Home (Part II)This post details the evolution of ML models for personalizing content on Spotify's Home feed. It covers the transition from a siloed experimentation platform to integrated Kubeflow clusters, highlighting challenges with custom evaluation metrics and comparing ML models against heuristic baselines. The post also discusses the development of custom dashboards for qualitative evaluation of recommendations and the struggles of automated model retraining and deployment, emphasizing the need for frequent retraining for certain models and the technical debt incurred by others. The authors contributed to the Python SDK for Kubeflow to support custom Python evaluations and developed a dashboard for manual evaluation of recommendations.
  5. For Your Ears Only: Personalizing Spotify Home with Machine LearningThis post details the application of Spotify's ML-driven personalization systems to the Spotify Home screen. It highlights the scale challenges compared to other platforms and the transition from custom data libraries to a standardized ML infrastructure using TensorFlow Extended (TFE) and Kubeflow. Key technical details include the use of a multi-armed bandit framework for balancing exploitation and exploration, counterfactual training, and the benefits of TFDV for data pipeline bug detection and consistency checks. The post emphasizes the shift towards engineer productivity and innovation by focusing on a 'Paved Road' approach to ML infrastructure.
  7. Introducing The Million Playlist Dataset and RecSys Challenge 2018This post introduces the Million Playlist Dataset (MPD), a dataset of 1 million user-created playlists sampled from Spotify's public playlists, containing over 2 million unique tracks by nearly 300,000 artists. It also announces the RecSys Challenge 2018, focused on automatic playlist continuation, where researchers are tasked with predicting hidden songs in real user-created playlists. The MPD includes playlist titles, track lists with metadata, and other anonymized metadata fields. The challenge aims to advance research in recommender systems for music.
  8. Commoditizing Music Machine Learning : ServicesThis post details the commoditization of music machine learning services by building a similarity infrastructure and a service-oriented architecture to manipulate vectors and build recommendations as a last-mile effort. It describes the development of an orchestration process ('doorman') to ensure atomic updates of musical vectors across the system, separating user vector generation from training and enabling real-time reactions to user behavior. The post also mentions the migration of infrastructure to Google Cloud and the challenges of building a consistent feedback framework across distributed models.
  9. What made Discover Weekly one of our most successful feature launches to date?This post details the genesis and early learnings of Discover Weekly, a highly successful personalized playlist feature. It highlights the iterative process of combining existing concepts ('Play it Forward' playlist from Year in Music and the 'Discover' page) to create a new user experience. Key technical and product insights include the importance of playlist simplicity and 'humanness,' the impact of weekly refreshes on habit formation, the role of personalized imagery in creating ownership, and the validation of underlying personalization technology through user feedback. It also emphasizes the crucial involvement of marketing and design teams early in the development process.
  10. Personalization at Spotify using CassandraThis post details the architecture and technical decisions behind Spotify's personalization system, specifically highlighting the use of Cassandra for storing user profile attributes and entity metadata. It explains the rationale for choosing Cassandra, focusing on its ability to scale horizontally, support cross-site replication, provide low-latency operations with tunable consistency, and facilitate bulk data transfer from Hadoop. The post also outlines the overall system architecture, including the roles of Kafka, Storm, and Crunch, and discusses the evolution of the Cassandra data model for entity metadata and user profiles.