Music Recommendations Based on Real-Time Data

Abstract

This thesis describes the development, implementation and results of a music recommender system that utilizes real time data, namely time and heart rate, for the recommendations. The recommender system was made by combining two systems, the recommender system which predicts a number of song features for a specific user and a ranking system which finds the best matching tracks for these features. Three implementations of the recommender system were implemented for comparison, namely Deep Neural Network, Contextual Bandit and Linear Regression. These implementations were tested with offline evaluation which showed that for our problem, a contextual bandit model had the best accuracy.