10.5281/zenodo.3257096
https://zenodo.org/records/3257096
oai:zenodo.org:3257096
Christos Koutlis
Christos Koutlis
CERTH ITI
Manos Schinas
Manos Schinas
CERTH ITI
Vasiliki Gkatziaki
Vasiliki Gkatziaki
CERTH ITI
Symeon Papadopoulos
Symeon Papadopoulos
CERTH ITI
Yiannis Kompatsiaris
Yiannis Kompatsiaris
CERTH ITI
Data-Driven Song Recognition Estimation Using Collective Memory Dynamics Models
Zenodo
2019
2019-06-26
10.5281/zenodo.3257095
https://zenodo.org/communities/futurepulse
Creative Commons Attribution 4.0 International
Cultural products such as music tracks intend to be appreciated and recognized by a portion of the audience. However, no matter how highly recognized a song might be at the beginning of its life, its recognition will inevitably and progressively decay. The mechanism that governs this decreasing trajectory could be modelled as a forgetting curve or a collective memory decay process. Here, we propose a composite model, termed T-REC, that involves chart data, YouTube views, Spotify popularity of tracks and forgetting curve dynamics with the purpose of estimating song recognition levels. We also present a comparative study, involving state-of-the-art and baseline models based on ground truth data from a survey that we conducted regarding the recognition level of 100 songs in Sweden. Our method is found to perform best among this ensemble of models. A remarkable finding of our study pertains to the role of the number of weeks a song remains in the charts, which is found to be a major factor for the accurate estimation of the song recognition level.