Automatic Music Recommendation Based on Acoustic Content and Implicit Listening Feedback

Rodrigo Carvalho Borges,Marcelo Gomes De Queiroz

doi:10.5216/mh.v18i1.53569

Abstract

   Recommending music automatically isn’t simply about finding songs similar to what a user is accustomed to listen, but also about suggesting potentially interesting pieces that bear no obvious relationships to a user listen- ing history. This work addresses the problem known as “cold start”, where new songs with no user listening history are added to an existing dataset, and proposes a probabilistic model for inference of users listening interest on newly added songs based on acoustic content and implicit listening feedback. Experiments using a dataset of selected Bra- zilian popular music show that the proposed method compares favorably to alternative statistical models.   

Highlights

Automatic music recommendation entered the mainstream in the last decades after a huge amount of digital media became available through online services (LOGAN, 2004) (ADOMAVICIUS; TUZHILIN, 2005)
Recall, F-measure and area under the receiver-operator curve (AROC) are expressed as means and standard deviations, and are presented for both methods (LR and Codeword Bernoulli Average (CBA)) as functions of the size K of the VQ Mel-Frequency Cepstrum Coefficients (MFCC) histograms
This article proposed a novel binary predictor for music recommendation based on a Codeword Bernoulli Average (CBA) model applied to binary implicit listening feedbacks from users of a media streaming application built on top of collection of Brazilian popular music dataset

Summary

Introduction

Automatic music recommendation entered the mainstream in the last decades after a huge amount of digital media became available through online services (LOGAN, 2004) (ADOMAVICIUS; TUZHILIN, 2005). MFCCs are widely used in MIR-related tasks, being considered a representation that captures relevant timbre-related aspects of audio signals These authors proposed a group of latent variables corresponding to music pseudo-genres from which the user would have to choose, and the system selected a song using a stochastic method. An experiment was conducted by collecting users implicit listening feedback through a media player app that continuously recommended new songs to users and recorded if they skipped or listened them to the end The goal of this experiment was to assess the predictive power of the CBA model to obtain the binary values corresponding to a user’s implicit feedback.

CBA model for music recommendation

Listening Data and Prediction Experiments

Results and Discussion

Conclusions and future work