Combining Deep Learning and Game Theory for Music Genre Classification

Abstract

In this thesis we have studied how a technique based on how game theory can improve classification results obtained with a deep learning module. In order to get this improvement we have applied this method to the music genre classification problem, comparing the obtained results. The proposed model is composed by a convolutional recurrent neural network (CRNN), that deals with classifying every single element, and the Graph Transduction Game (GTG) method, that allows to compare these elements on the basis of a similarity measure and thus exploit the contextual information in order to get a better classification. The idea behind this work is that the neural network architecture does not directly exploit the information coming from the comparison of the observations passed as input. Therefore we think that the introduction of a module in charge for this purpose can improve final accuracy, especially when we work with limited datasets. In order to assess the effect of the proposed approach we have performed experiments on benchmark datasets and we report the results obtained.