[3Xin4] Poster session 1

[3Xin4-08] Consideration on Characteristics of Music Embeddings from Playlist

〇Yoshihiro Morii¹, Hiroki Shibata¹, Yasufumi Takama¹ (1.Tokyo Metropolitan University)

Keywords:distributed representation, playlist, music embeddings

This paper investigates the characteristics of music vectors obtained from music playlists using word2vec and Paragraph Vector. It is known that a distributed representation such as obtained by word2vec and Paragraph Vector has the property that vectors of similar data are located close to each other in a vector space. Vector operations between words have a property called Additive Compositionality, which reflects the meaning of words. A previous study has reported a distributed representation of songs can be obtained from music playlists and the similarity of songs by the same artist was high. However, it has not investigated whether Additive Compositionality holds or not.This paper focuses on the artist and season as elements related to songs, and performed vector operations in terms of those elements to verify Additive Compositionality. Experimental results show that cases where Additive Compositionality holds were confirmed.