This project focuses on Spotify, the most popular audio-streaming service with over 489 million users. One of the key Spotify advantage is recommendations systems, which are pretty lovable by audience and very intimidating to competitors. These recommendation systems are based on advanced models built on different types of data, including raw audio features. Spotify is keeping and updating 13 such features for more than 100 millions of songs. It is clear that maintaining even 1 such feature for such a huge amount of songs cost a round sum. Also, some of these features can correlate and due to multicollinearity create a risk of harm to the models. In this context, question arise: are all these 13 features needed indeed? Cannot answer without PCA!

Mini-research made within internship in Sociology of Education and Science Laboratory, HSE University

Collection of projects completed by the Third Temple Builders team: Dina Amaltdinova, Nikita Sharin, Ekaterina Teterina, Ekaterina Drobyshevska.