Comparative Music Similarity Modelling Using Transfer Learning Across User Groups

We introduce a new application of transfer learning for training and comparing music similarity models based on relative user data: The proposed Relative Information-Theoretic Metric Learning (RITML) algorithm adapts a Mahalanobis distance using an iterative application of the ITML algorithm, thereby extending it to relative similarity data. RITML supports transfer learning by training models with respect to a given template model that can provide prior information for regularisation. With this feature we use information from larger datasets to build better models for more specific datasets, such as user groups from different cultures or of different age. We then evaluate what model parameters, in this case acoustic features, are relevant for the specific models when compared to the general user data. We to this end introduce the new CASimIR dataset, the first openly available relative similarity dataset with user attributes. With two age-related subsets, we show that transfer learning with RITML leads to better age-specific models. RITML here improves learning on small datasets. Using the larger MagnaTagATune dataset, we show that RITML performs as well as state-of-the-art algorithms in terms of general similarity estimation.