Analysis of Song/Artist Latent Features and Its Application for Song Search
Analysis of Song/Artist Latent Features and Its Application for Song Search
复制标题
DOI：
10.5281/zenodo.4245538
复制
发表时间：
2020-10
期刊：
影响因子：
--
通讯作者：
Kosetsu Tsukuda;Masataka Goto
中科院分区：
文献类型：
--
作者：
Kosetsu Tsukuda;Masataka Goto
为了向用户推荐歌曲，一种有效的方法是用潜在向量表示艺术家和歌曲并预测用户对歌曲的偏好。尽管潜在向量很好地代表了艺术家和歌曲的特征，但它们通常仅用于计算偏好分数。在本文中，我们讨论如何利用这些向量来实现使用户能够从新的角度搜索歌曲的应用程序。为此，通过将歌曲/艺术家向量嵌入到同一特征空间中，我们首先提出艺术家-歌曲关系的两个概念：整体相似性和突出亲和力。总体相似度是指歌曲的特征与歌手的特征总体相似的程度；突出亲和力是指歌曲突出地体现艺术家特征的程度。通过使用Last.fm两年的播放日志，我们分析了概念的特征。此外，根据分析结果，我们提出了三种歌曲搜索应用。通过案例研究，我们证明我们提出的应用程序有利于根据用户的各种搜索意图搜索歌曲。
For recommending songs to a user, one effective approach is to represent artists and songs with latent vectors and predict the user's preference toward the songs. Although the latent vectors represent the characteristics of artists and songs well, they have typically been used only for computing the preference score. In this paper, we discuss how we can leverage these vectors for realizing applications that enable users to search for songs from new perspectives. To this end, by embedding song/artist vectors into the same feature space, we first propose two concepts of artist-song relationships: overall similarity and prominent affinity. Overall similarity is the degree to which the characteristics of a song are similar overall to the characteristics of the artist; while prominent affinity is the degree to which a song prominently represents the characteristics of the artist. By using Last.fm play logs for two years, we analyze the characteristics of the concepts. Moreover, based on the analysis results, we propose three applications for song search. Through case studies, we demonstrate that our proposed applications are beneficial for searching for songs according to the users' various search intents.