Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Advancing Cultural Inclusivity: Optimizing Embedding Spaces for Balanced Music Recommendations

Armin Moradi, Nicola Neophytou, Golnoosh Farnadi

TL;DR

The paper tackles cultural and popularity bias in music recommendations by analyzing the embedding-space geometry of a prototype-based matrix factorization method (ProtoMF). It introduces two in-space enhancements—Prototype-K-filtering and Prototype-Distributing Regularizer—that modify how prototypes are used and distributed without changing the original loss function. Empirical results on LastFM-2b and MovieLens-1M demonstrate reduced disparities between underrepresented and overrepresented groups while maintaining or improving recommendation quality. This work advances fairer, more inclusive music discovery and highlights the potential of embedding-space design for transparent, culturally aware recommendations.

Abstract

Popularity bias in music recommendation systems -- where artists and tracks with the highest listen counts are recommended more often -- can also propagate biases along demographic and cultural axes. In this work, we identify these biases in recommendations for artists from underrepresented cultural groups in prototype-based matrix factorization methods. Unlike traditional matrix factorization methods, prototype-based approaches are interpretable. This allows us to directly link the observed bias in recommendations for minority artists (the effect) to specific properties of the embedding space (the cause). We mitigate popularity bias in music recommendation through capturing both users' and songs' cultural nuances in the embedding space. To address these challenges while maintaining recommendation quality, we propose two novel enhancements to the embedding space: i) we propose an approach to filter-out the irrelevant prototypes used to represent each user and item to improve generalizability, and ii) we introduce regularization techniques to reinforce a more uniform distribution of prototypes within the embedding space. Our results demonstrate significant improvements in reducing popularity bias and enhancing demographic and cultural fairness in music recommendations while achieving competitive -- if not better -- overall performance.

Advancing Cultural Inclusivity: Optimizing Embedding Spaces for Balanced Music Recommendations

TL;DR

The paper tackles cultural and popularity bias in music recommendations by analyzing the embedding-space geometry of a prototype-based matrix factorization method (ProtoMF). It introduces two in-space enhancements—Prototype-K-filtering and Prototype-Distributing Regularizer—that modify how prototypes are used and distributed without changing the original loss function. Empirical results on LastFM-2b and MovieLens-1M demonstrate reduced disparities between underrepresented and overrepresented groups while maintaining or improving recommendation quality. This work advances fairer, more inclusive music discovery and highlights the potential of embedding-space design for transparent, culturally aware recommendations.

Abstract

Popularity bias in music recommendation systems -- where artists and tracks with the highest listen counts are recommended more often -- can also propagate biases along demographic and cultural axes. In this work, we identify these biases in recommendations for artists from underrepresented cultural groups in prototype-based matrix factorization methods. Unlike traditional matrix factorization methods, prototype-based approaches are interpretable. This allows us to directly link the observed bias in recommendations for minority artists (the effect) to specific properties of the embedding space (the cause). We mitigate popularity bias in music recommendation through capturing both users' and songs' cultural nuances in the embedding space. To address these challenges while maintaining recommendation quality, we propose two novel enhancements to the embedding space: i) we propose an approach to filter-out the irrelevant prototypes used to represent each user and item to improve generalizability, and ii) we introduce regularization techniques to reinforce a more uniform distribution of prototypes within the embedding space. Our results demonstrate significant improvements in reducing popularity bias and enhancing demographic and cultural fairness in music recommendations while achieving competitive -- if not better -- overall performance.
Paper Structure (16 sections, 5 equations, 3 figures, 2 tables)

This paper contains 16 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Prototype-based Matrix Factorization
  5. Methodology
  6. Prototype K-filtering
  7. Prototype-Distributing Regularizer
  8. Empirical Evaluation
  9. Datasets
  10. Baselines
  11. Fairness Metrics
  12. RQ1: Do prototypes in the embedding space of ProtoMF possess inherent properties that can lead to cultural bias?
  13. RQ2: How successful are our proposed approaches, which modify the learning process of the embedding space to enhance the discoverability of minority cultures and mitigate popularity bias, compared to existing methods?
  14. Conclusion & Future Directions
  15. Ethical Considerations and Broader Impacts
  16. ...and 1 more sections

Figures (3)

  • Figure 1: t-SNE visualizations of the item embedding space for (a) the baseline ProtoMF model and (b) the model after applying both Prototype k-Filtering and the Prototype-Distributing Regularizer. Larger dots represent more popular items. The triangles indicate the positions of prototypes. NL and US represent an underrepresented and overrepresented country, respectively.
  • Figure 2: Average distance of items to k nearest prototypes, grouped by popularity, for the baseline ProtoMF method. The gap in average distances between items of different popularity levels narrows with smaller k values.
  • Figure 3: Average rank of overrepresented items (from US, GB, CA) and underrepresented items (from BR, NL, PL), comparing ProtoMF, ZeroSum and our model using $k$ and $\lambda$.