Probabilistic approaches for music similarity using restricted Boltzmann machines

In music informatics, there has been increasing attention to relative similarity as it plays a central role in music retrieval, recommendation, and musicology. Most approaches for relative similarity are based on distance metric learning, in which similarity relationship is modelled by a parameterised distance function. Normally, these parameters can be learned by solving a constrained optimisation problem using kernel-based methods. In this paper, we study the use of restricted Boltzmann machines (RBMs) in similarity modelling. We take advantage of RBM as a probabilistic neural network to assign a true hypothesis “x is more similar to y than to z” with a higher probability. Such model can be trained by maximising the true hypotheses while, at the same time, minimising the false hypotheses using a stochastic method. Alternatively, we show that learning similarity relations can be done deterministically by minimising the free energy function of a bipolar RBM or using a classification approach. In the experiments, we evaluate our proposed approaches on music scripts extracted from MagnaTagATune dataset. The results show that an energy-based optimisation approach with bipolar RBM can achieve better performance than other methods, including support vector machine and machine learning rank which are the state-of-the-art for this dataset.