Analysis of Data Models for Recommender Systems using Factorization Machines

With the rise of music streaming platforms like Spotify, the field of music recommender systems is an ever-growing field of research. Contextual information is often used to further increase the performance of a recommender system. In this thesis, we investigate different contextual features and analyze their usefulness in the field of music recommender systems. We further utilize Factorization Machines, a state of the art machine learning algorithm, as a predictor. To find feature combinations, a dataset with contextual information is created from an existing playlist dataset. A grid search is conducted to find the best performing combination of features and hyperparameters for the Factorization Machines' different learning algorithms. Since the amount of combinations of features and parameters is quite extensive, the features are grouped into user, track and artist features to reduce the dimensions of the grid. For evaluation, an experimental framework is created. The results show that models only containing user features perform the best in most of the tested cases.