MUSIC RECOMMENDATION SYSTEM

Recommender SystemCollaborative FilteringContent-based Recommendations

CONTEXT AND OBJECTIVE

Streaming platforms have become crucial in modern society, providing instant access to extensive libraries of movies, music, podcasts, and videos. These services depend on consumer engagement time for revenue, but users can be overwhelmed by the vast amount of available content.

Recommendation Systems (RS) solve this by using advanced algorithms and data to offer personalized suggestions, enhancing user experience and engagement. For instance, Netflix attributes 80% of its viewing hours to personalized recommendations, while Spotify relies on AI to create custom playlists, with 30% of its streams coming from AI-recommended songs.

In the framework of the MIT Applied Data Science Program, the goal of this Capstone project was to develop a recommendation system for a global audio content provider offering a 1 million songs catalog to its 75,000+ customers.

WHAT WAS DONE

An hybrid RS built on collaborative filtering and content-based recommandations was proposed. With an RMSE of 0.94 when predicting the ratings (from 1- to 5-stars) of songs that haven’t been consumed yet, this system is expected to bring business value by increasing the average engagement time by 10%. It provides suggestions personalized to each user for the next songs to play, therefore improving their experience.

The core component of this hybrid system uses a matrix factorization engine that serves regular users, while a complentary content-based engine helps new users get started on the streaming platform. Based on a computationally efficient design, the solution is scalable and can provide real-time suggestions even as the customer base grows.

Finding the right model(s)

Several approaches were tried to solve the business problem. A popularity algorithm simply suggests the most popular songs to all users, regardless of their personal tastes. The content approach looks for similar titles by exploiting the songs metadata (title, artist, etc.) Collaborative filtering is a family of models that takes advantage of similarties between users and songs to make personalized suggestions. These models provided the best results.

The hybrid model

A matrix factorization model was proposed as the core component of the recommendation system. The suggestions it makes are relevant to users 45% of time, but to work properly users need to have listened to at least 50 titles. To adress this so-called ‘cold start’ problem, a content-based engine was added to offer a solution to new users.

Recommendations example

These are the top-5 recommendations for a specific user, based on the predicted ratings.