Towards a better understanding of music playlist titles and descriptions

Hao, Yun

Permalink

https://hdl.handle.net/2142/110509 Copy

Description

Title

Towards a better understanding of music playlist titles and descriptions

Author(s)

Hao, Yun

Issue Date

2021-04-20

Director of Research (if dissertation) or Advisor (if thesis)

Downie, J. Stephen

Doctoral Committee Chair(s)

Downie, J. Stephen

Committee Member(s)

• Hu, Xiao
• Torvik, Vetle
• Ehmann, Andreas
• Bosch, Nigel

Department of Study

Information Sciences

Discipline

Information Sciences

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• natural language processing
• recommender systems

Abstract

Music playlists, either user-generated or curated by music streaming services, often come with titles and descriptions. Although informative, these titles and descriptions make up a sparse and noisy semantic space that is challenging to be leveraged for tasks such as making music recommendations. This dissertation is dedicated to developing a better understanding of playlist titles and descriptions by leveraging track sequences in playlists. Specifically, work has been done to capture latent patterns in tracks by an embedding approach, and the latent patterns are found to be well aligned with the organizing principles of mixtapes identified more than a decade ago. The effectiveness of the latent patterns is evaluated by the task of generating descriptive keywords/tags for playlists given tracks, indicating that the latent patterns learned from tracks in playlists are able to provide a good understanding of playlist titles and descriptions. The identified latent patterns are further leveraged to improve model performance on the task of predicting missing tracks given playlist titles and descriptions. Experimental results show that the proposed models yield improvements to the task, especially when playlist descriptions are provided as model input in addition to titles. The main contributions of this work include (1) providing a better solution to dealing with ``cold-start'' playlists in music recommender systems, and (2) proposing an effective approach to automatically generating descriptive keywords/tags for playlists using track sequences.

Graduation Semester

2021-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/110509

Copyright and License Information

Copyright 2021 Yun Hao

Owning Collections