Streaming Data Reveals UK's Regional Music Tastes
An analysis of Spotify and YouTube streaming data has mapped the distinct regional music preferences across Britain. The data reveals significant differences in taste by city, such as the popularity of Oasis in Scotland and Sam Fender in Newcastle. These findings underscore the need for recommendation systems to incorporate geo-segmentation for better user engagement.
- A year-long analysis of Spotify's Top 200 charts in 73 countries revealed that American artists occupy 55% of the UK's chart, nearly double the 29% share of British artists, ranking the UK 39th out of 73 countries in listenership of its own domestic talent. - Netflix's recommendation system addresses regional differences by moving from separate, country-specific models to a single global model. This approach prevents larger countries from dominating regional models and allows the system to identify worldwide taste communities, improving recommendations for niche interests. - To productionalize and validate geo-segmented features, firms use geo-based A/B testing, which treats entire geographic regions as control or test groups. Instacart’s engineering team implements this by randomizing delivery zones to measure user metrics, comparing treatment and control zones with a Difference-in-Difference model to minimize engineering overhead. - Spotify’s architecture separates its high-availability, low-latency personalization pipelines from its experimentation systems. This allows for constant testing of new ranking logic and recommendation models without risking the stability of the live, user-facing product. - The core of many large-scale recommendation systems at FAANG companies involves pre-computing features offline from user interaction data. These features, often simple counters and rates (e.g., number of times a user has clicked on a product in a category), form the bedrock of even sophisticated ranking models. - Spotify is now leveraging Large Language Models (LLMs) to move beyond simple recommendations and create "personal narratives" that explain why a track is suggested. By fine-tuning models like Llama on curated internal data, they can generate contextual explanations that help users discover new artists. - While collaborative filtering remains a key technique, platforms face challenges with it, such as data sparsity and scalability. Netflix is developing a foundation model for recommendations, inspired by LLMs in NLP, to centralize member preference learning from comprehensive interaction histories, allowing innovations to be transferred across different recommendation models more easily.
