Um modelo para classificação de músicas voltado à plataforma de streaming utilizando aprendizado de métricas e predição estruturada

Abstract

In the scenario where decision making can change the behavior of next actions, an algorithm requires a lot of adaptive and cognitive power. For example, we can highlight choosing a route to go to work, buying a product via e-commerce, investing in the financial market, and selecting a playlist using a streaming platform. Developing models for learning similarity between pieces of music and extracting features from audio media files is an increasingly important task for the entertainment industry due to the difficulties of obtaining consistent metadata information. This paper proposes a new model of music classification based on metric learning, feature extraction from MP3 audio files and dimensionality reduction. The metric learning process considers the learning of a set of parameterized distances employing a structured prediction approach from a set of pieces of music distributed over various musical genres. The main objective of this paper is to enable the learning of a custom metric for each client that was reached through two approaches, offline and online. For the offline solution, we attest to the model’s validity by conducting a set of experiments and comparing training and test results with baseline algorithms such as K-means and SVM. The experiments showed promising results and encouraged the development of an online version of the learning model. Thereby, we made the online solution that aimed at real-time learning. Through a set of experiments, we made an evaluation and compared to results obtained with the offline version. Experiments show that the online version accuracy converges to the ideal offline solution, minimizing the average regret. The performance of the learning model makes it possible to extend work to a variety of applications beyond the music classification.