Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

User-centric Music Recommendations

Jaime Ramirez Castillo, M. Julia Flores, Ann E. Nicholson

TL;DR

The paper tackles the lack of explainability in music recommendations by proposing a user-centric, explainable pipeline that personalizes recommendations for moments in time using a single user’s historical Last.fm plays and Spotify features. It introduces a four-phase framework that first estimates moment-specific tag strengths, then predicts target Spotify audio features (danceability in the study), and finally ranks tracks by similarity to the predicted features, with Phase Four planned for diversity and user control. A musical moments dataset is created by aggregating 90k+ plays into 14,203 hourly moments with 1,000 tag features and 12 Spotify features; phase-two regression models are evaluated, with XGBoost providing the best RMSE (0.091) compared to baselines. The work demonstrates a path toward interpretable, time-aware personalized music recommendations and outlines clear directions for extending to multiple users, multi-target targets, and richer explainability techniques.

Abstract

This work presents a user-centric recommendation framework, designed as a pipeline with four distinct, connected, and customizable phases. These phases are intended to improve explainability and boost user engagement. We have collected the historical Last.fm track playback records of a single user over approximately 15 years. The collected dataset includes more than 90,000 playbacks and approximately 14,000 unique tracks. From track playback records, we have created a dataset of user temporal contexts (each row is a specific moment when the user listened to certain music descriptors). As music descriptors, we have used community-contributed Last.fm tags and Spotify audio features. They represent the music that, throughout years, the user has been listening to. Next, given the most relevant Last.fm tags of a moment (e.g. the hour of the day), we predict the Spotify audio features that best fit the user preferences in that particular moment. Finally, we use the predicted audio features to find tracks similar to these features. The final aim is to recommend (and discover) tracks that the user may feel like listening to at a particular moment. For our initial study case, we have chosen to predict only a single audio feature target: danceability. The framework, however, allows to include more target variables. The ability to learn the musical habits from a single user can be quite powerful, and this framework could be extended to other users.

User-centric Music Recommendations

TL;DR

The paper tackles the lack of explainability in music recommendations by proposing a user-centric, explainable pipeline that personalizes recommendations for moments in time using a single user’s historical Last.fm plays and Spotify features. It introduces a four-phase framework that first estimates moment-specific tag strengths, then predicts target Spotify audio features (danceability in the study), and finally ranks tracks by similarity to the predicted features, with Phase Four planned for diversity and user control. A musical moments dataset is created by aggregating 90k+ plays into 14,203 hourly moments with 1,000 tag features and 12 Spotify features; phase-two regression models are evaluated, with XGBoost providing the best RMSE (0.091) compared to baselines. The work demonstrates a path toward interpretable, time-aware personalized music recommendations and outlines clear directions for extending to multiple users, multi-target targets, and richer explainability techniques.

Abstract

This work presents a user-centric recommendation framework, designed as a pipeline with four distinct, connected, and customizable phases. These phases are intended to improve explainability and boost user engagement. We have collected the historical Last.fm track playback records of a single user over approximately 15 years. The collected dataset includes more than 90,000 playbacks and approximately 14,000 unique tracks. From track playback records, we have created a dataset of user temporal contexts (each row is a specific moment when the user listened to certain music descriptors). As music descriptors, we have used community-contributed Last.fm tags and Spotify audio features. They represent the music that, throughout years, the user has been listening to. Next, given the most relevant Last.fm tags of a moment (e.g. the hour of the day), we predict the Spotify audio features that best fit the user preferences in that particular moment. Finally, we use the predicted audio features to find tracks similar to these features. The final aim is to recommend (and discover) tracks that the user may feel like listening to at a particular moment. For our initial study case, we have chosen to predict only a single audio feature target: danceability. The framework, however, allows to include more target variables. The ability to learn the musical habits from a single user can be quite powerful, and this framework could be extended to other users.
Paper Structure (26 sections, 4 figures, 2 tables)

This paper contains 26 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. A User-centric Recommendation Framework
  3. A Single-user Dataset
  4. User Listening Context Data
  5. Track Characterization Data
  6. The Phases of the Recommendation Framework
  7. Methodology
  8. Data Gathering
  9. Last.fm Tags
  10. Spotify Audio Features
  11. Data Preparation
  12. Last.fm Tags Reduction
  13. Reducing the Number of Samples
  14. Spotify Features Reduction
  15. The Resulting Musical-moments Dataset
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Data gathering. We have collected the historical logs of a single user from Last.fm, for the 2007-2021 period. The data includes more than 90,000 track playback records. For each track the user has ever listened to, we have gathered Spotify audio features.
  • Figure 2: Data preparation.
  • Figure 3: Distribution of the danceability Spotify audio feature across 20,000 individual tracks, which correspond to the single-user dataset used in this paper.
  • Figure 4: Recommendation pipeline. The recommendation pipeline begins with a particular moment in time (the hour of the day). Phase one aggregates the available Last.fm tag strength values for this particular hour. Phase two predicts the most-suitable Spotify audio features for this hour, and finally Phase three selects the tracks that are closest to this prediction. The diagram also depicts Phase Four, to provide an idea of where this additional phase fits in the framework.