Feature Selection for Composer Classification Method using Quantity of Information
Ayaka Takamoto, Mitsuo Yoshida, Kyoji Umemura, Yuko Ichikawa · 2018
This paper presents a method to select music features and demonstrates how long a music phrases should be in order to be use as a feature of a composer classification task using the quantity of information. Obtaining the quantity of information is equivalent to estimating the likelihood; or how likely the music piece appears assuming the piece is drawn from set of pieces of the selected composer. We compute the likelihood as the product of the probability of phrases that form the piece. Although it might be natural to consider all possible partition of phrase, we propose limiting the length of phrases to be considered, and obtain the minimum value of the required length of the phrase. We find that the phrases within only one transition are highly effective.