Music similarity analysis using the big data framework spark

Abstract

A parameterizable recommender system based on the Big Data processing framework Spark is introduced, which takes multiple tonal properties of music into account and is capable of recommending music based on a user's personal preferences. The implemented system is fully scalable; more songs can be added to the dataset, the cluster size can be increased, and the possibility to add different kinds of audio features and more state-of-the-art similarity measurements is given. This thesis also deals with the extraction of the required audio features in parallel on a computer cluster. The extracted features are then processed by the Spark based recommender system, and song recommendations for a dataset consisting of approximately 114000 songs are retrieved in less than 12 seconds on a 16 node Spark cluster, combining eight different audio feature types and similarity measurements.